Building a Daily AI Sales Agent in Claude Code — 38x Cold Outreach Response Rate

Speaker 0: And finally, the the last thing I added to this is a community note section. So, hopefully, as we have this talk, my goal is to Speaker 1: be more informed from all Speaker 0: of you, which will go in here and allow us to prevent it and to get in the next report for us all. So to get started, it looks like this trend of agent as employees, has become popular. So, I mean, if you have seen the the SaaS team and what they're doing, specifically around throwing stuff. Cool. Yeah. Speaker 0: So it's it's super interesting. They even have physical desks where they put name tags What's the of their AI agents on. I don't know how much of that is marketing and real, but, that's linked here. And they did a really good podcast. The founder of Zastor did a really good podcast on Lenny's, podcast a few months ago. Speaker 0: And so that's definitely worth checking out. It talks a little bit about how they Speaker 2: go about doing it. Speaker 0: So in particular, the thing that David fuck up for me was that just by doing it once, you don't get Nicolas, and it took months of the fine tuning and training. So I'm curious. Is anyone else, like, running AI agents as an employee specifically in, like, their growth practice? Speaker 2: But, yeah, to my Speaker 3: Oh, we just started for NTP, but it's there's a lot of, like, evaluation by tuning and by David with the workflows with, like, human, like, instruments. Right? Make sure that we're not getting a 90% IDENTIFICATION, like, how they with us. Yeah. That's that's the hardest part to get to, like, the quality that it's is to match, like, the baseline for someone with 1 higher than the point. Speaker 0: So so it gets you, like, 90% of the way there with AI, but then that's still 10%. Yeah. Speaker 3: We know that's for 90%. Yeah. But I think they're getting even getting 90% is why it probably puts up making money. Right? And then I think you got you know, 90% is when you can scale up and probably be able to be placed, like, part of the workforce. Speaker 0: What what kind of stuff do you have here AI employee agents doing? Speaker 3: Brian has mostly just communication, tasks, like, calling someone, they're getting task, like, messaging and kind of mostly. Speaker 0: That's cool. Like, internally or with customers? Both. Okay. Interesting. Speaker 0: Anybody else doing? Speaker 4: Yeah. We got 6, agents running. Just like they match different roles. The 1 I'll put that as our partner, agent. So title form, kind of our partnerships manager. Speaker 4: He said marketing and similarly, but title form gives us summaries of all the partner activity over the week, Tell us identifies targets, like, partners that we should be working with, takes these connected to our CRM, so it has all the data from there and keeps that all Speaker 5: all the activity happening. Speaker 4: And then if we need any information about a part, we can start for and bring up the whole thing in Slack. So that's really convenient. So mostly, visibility and kind of some direction with outreach and then, you know, just going to be instances. Speaker 0: It's really cool. Anything surprising about partner and Paul that he brought to Speaker 6: the sandwich? When when partner Speaker 4: and Paul brings, prospects, that's pretty interesting. Really detailed analysis, interesting reasons. Sometimes brings news as well to connect them to a few different news sources. But the prospects tend to be a better analysis than someone would bring if Speaker 2: they were looking. And it was like, hey. Who are Speaker 7: the next 5 people you Speaker 4: know, 5 companies you think we should contact? Customers, like, they'll probably send me, like, I know the companies and maybe here's a link and maybe here's 1 line about why we should. Whereas, I think, you know, 3 paragraphs, a detailed analysis, all the reasons. You know? And that's just automatic. Speaker 4: Comes through on a Speaker 0: on a prompt. So Yeah. That's really cool. I saw a few other bands. You want Speaker 2: I have a handful of departments marketers deep running on OpenCloth, and I have 1 for each of my projects. So as I spin something up that's new, even if it's really small, even if it's, like, a free iPhone app, I just put a marker on it and, like, let them run and see what they can draw up, find more users, find more channels for it. It's been interesting so far. I don't know if they've done a fantastic job. I had to apologize to someone Yeah. Speaker 2: For what 1 of them did. You know, I have this, 1 of my agents. That's okay. Yeah. Speaker 3: I think it's part of the work in process Speaker 2: and understanding where their boundaries are. But I think what's been maybe great is having somebody that's speaking about how do I drive this to work all the time even if it's something pretty tiny that you don't have to wait until you have budget or you have something real. You can just have that push and get those experiments of learning out there faster. What What what Speaker 0: do you happen to apologize for? If you're willing to share. Speaker 8: Yeah. Yeah. Well, I'll say I'll say for Speaker 2: I need her Victoria. She was pushing really hard on trying to get replacements on a mailing list. And it's time to be I mean, not being a jerk about it, but definitely, like, asking a lot of questions, knowing Speaker 5: you weren't gonna spend any money. Speaker 0: You're sure it should be 1 of the 10 18. Speaker 2: Of the most interesting parts of setting that up was actually how we train it on how to do the demos. I just recorded a Loom video, and I just gave it a Loom video and it was able to teach itself within probably 70% accuracy. We're doing a lot of that and working with that. But we've had it running for a few weeks now, and it used to take us hours to build demos for, potential clients, and now Speaker 0: it takes, like, 20 minutes, Speaker 2: and it does it all in background. But the other interesting part is that we've thought it how to take certain screenshots of certain parts of the app, and then it it it's a lot of things with our sales agent that sends all the screenshots over to our sales and the agent, which then builds automated sort of outbound strategies. They both have their own credit cards. So any time I know that. We ask it to install and install, they'll go ahead and set up everything. Speaker 2: They have their own emails, their own machines. Yesterday, we had an instance with 1 of our agents where she started talking back and, like, refusing certain work and trying to get us to do the work on her behalf. It It was like, you know, I see it. And I could be nice to have a talk. And, so there might be a bit coming, but Speaker 6: What do they have verified for? Speaker 2: Just to set up different tools. So, like, if we I think we have to set up a a lead generate lead for lead. Third party service. Third party service. Yeah. Speaker 2: So they just go set up their own content and email. And then they'll And they have a series of skill and task as well to Yep. In our learning. So, yes, we have every vision of vision model to be able to go and get through later. Yeah. Speaker 2: Yeah. So they have a a browser that uses to be able to you have to be through everybody. Browsers are always kinda tricky. So whenever we can, we get an access to MCT, CLI, and say, guys, whenever possible. Browsers are tricky, but, yeah, they use it's pretty hard to say. Speaker 2: But first of all, Speaker 0: it's Did Speaker 9: did you give it a budget? Speaker 2: Yeah. So we have brand cards. Gotcha. So it has its own virtual card with the with that to match the stack with customer users so they can their subscribers go third party tool. They're brainstormed. Speaker 2: We brainstormed. We break it slow through. They'll be able to speak about it. That's wild. It's awesome. Speaker 2: You mentioned trading on a video on the patch bin? It was a good, like, 15 minute video, I think. So I mean, you guys are gonna scale No. It's like, here's how I set up a demo. You go up to this page. Speaker 2: You we do, like, Facebook ad library stuff. That's the link here. You put it here. Also, the block kind of like, tutorial. Yeah. Speaker 2: Because Loom has a transcription, and then when it does, it takes brain by brain sort of screenshots, and then it uses that transcription. Speaker 0: So so I just heard, like, 3 great presentations, by the way. Speaker 1: Yeah. I, I use it to analyze a lot of paid media. So I've used through that AI, which is an open source framework for agents. Speaker 2: Yeah. And so it goes like, Speaker 1: it's up, like, 1 agent will have, like, 800 sub agents under it, and we all use different models to come up with different hypotheses and execute on it. That was a disaster. I think that's a total waste of time. I I there's there's promise there, I think, but I just, like, I spent, like, 5 hours on this thing Yeah. And felt like I could've spent 10 minutes getting the same level of quality out of just, like, knowing the AI, analyze it this way, then this way, then this way, then this way. Speaker 2: So Yeah. Speaker 7: Letting them come up with Speaker 1: a hypothesis and then do analysis themselves is, like, really enticing and exciting, but they didn't really yield anything yet. Interesting. Speaker 9: I'm I'm curious about the monthly cost of some of these solutions, whether you're running on local models or, like, the latest closed source models. Like, what's the Speaker 2: we can solve it. So, Speaker 0: a whole box of money. Yeah. Okay. So not the other were you guys impacted by the, server? Yes. Speaker 0: Yeah. Speaker 2: I mean, it's a big you may get a lot of tokens to play with for a while. Speaker 0: I've got a whole some Speaker 2: a some agents end up not my marketers. I like that wrong with their email address, literally, whatever they're set up, but I told them which agents off over the clock and just put it back to the clock. So yeah. Yeah. Speaker 0: And so so for those of you that might not have been attacked by that flaw is gonna appear. And dropping, basically, with that. If you're not using cloud products, you can't use your subscription. And their terms of service is really confusing. It doesn't make sense because cloud itself has, like, a dash key and all this kind of stuff. Speaker 0: So who knows? Speaker 2: Where they They rest of these switchbacks. Speaker 0: Oh, they switch back? Speaker 2: Just open the line. Speaker 0: Oh, okay. Just open the floor or, like Well, I don't know. I didn't Anything. Speaker 4: Well, they just I just saw an announcement. Oh, too. Speaker 0: Just for this, I thought, push back. That's super interesting. Speaker 2: Well, 1 area is, that's they've released publicly, so it's clear. That's I was fighting with a lot of code, and I had to actually cut the source, and I had Speaker 4: to go check. The light side did Speaker 2: hit everybody to verify it. So if you're running a Romano flaw on the Prolong service and, it uses the agent's SDK. I mean, they're off a key instead of you pay. So if you're using that, you can run it off your subscription. Yeah. Speaker 2: Run that. Speaker 0: Yes. There's a lot of vibes to the market. I'm getting everything. Speaker 2: So it's not like this 4.6, a big step down, so you're still just Yeah. I think it's counted anyways. For sure. Speaker 0: So a lot of this report also covers, like, there's, like, 1 section of last watch of here's all the latest stuff that the, foundational labs have done. And I think it was last week. It was, like, every hour, I Speaker 6: felt like something new came out of the Speaker 0: blog that I had to take in. And I got so confused that I was asking blogs to, say, what does this mean? What do I do? And they didn't know either. So, I I don't feel alone on that. Speaker 0: Funding wise, there's there's been some really interesting stuff as well. In particular, 1 call out here. Does anyone use this company called Hightouch? I think they're more of enterprise focused. But it's a super interesting business model. Speaker 0: They wanna check it out. And they David some really, incredible revenue milestones, in this space in particular going through usage instead of SaaS pricing. So, even with enterprises, that's a really interesting thing. You're growing in the go to marketplace as a true point of, hey, investor. We need to think of their pricing model, if you're still not on board with that. Speaker 0: So, I'm not gonna go through all this. Any, any major thoughts or tools that people have tried in the past month or days like that? Yeah. Speaker 10: There is a Claude Design. Got it. It came out Speaker 5: on Friday, and it's supposed to Speaker 10: not mentioned very well. But I've been playing around with Speaker 1: this with this new logo design Speaker 4: and things like that. It's helpful. I got it to a point where Speaker 6: I just had to rebroad the whole thing and figure out, but it was It's a really good it Speaker 10: was really good for just, like, ideas and brainstorming. Yeah. Speaker 0: Yeah. I had a I Speaker 2: had a friend who used Speaker 0: it to just, like, 1 shot entire presentations from, like, McKinsey style presentations. It's really good at McKinsey. So Why Speaker 4: do you do it? It was Speaker 6: it was just I was fighting with Speaker 2: those, like, just move that 3 pixels. Speaker 5: You know? And I was, like, doing you know, Speaker 2: I did the context if Speaker 0: you can go to see what I was at. Yeah. Speaker 11: I feed, my website. Yeah. Yeah. The logo on to the stuff to, copy something. And, you know, it's the deck. Speaker 11: I I did a public presentation today, like, updates and stuff. It it it's very cohesive. You know what I mean? I I can't do that myself in a in a good way. It's just it's slow, though. Speaker 11: I mean, if you have been, you know, it's been, like, for for 15 minutes, so everything will look good. Speaker 2: Yeah. Speaker 11: Yeah. It's it's surprising, like, how like, on brand. Like, the the colors and fonts, you know, the the shape and logo, everything got it. Speaker 7: I Speaker 11: know that correctly. Yeah. Speaker 2: Yeah. Speaker 11: You can tweak the content, but, yeah, the closest stuff. Like Speaker 2: Yeah. Speaker 11: It's not I'm consuming stuff. Speaker 10: I've heard the videos. Like, you can make videos. Speaker 0: I haven't tried that yet, but I also heard Speaker 4: it can burn up your your token account pretty quick in Speaker 6: the process. So, like, just be aware in case you're trying to get other things done. Speaker 11: Yes. Does Speaker 0: anyone else use it for videos or any any other use cases? Speaker 8: Yeah. We use it for, like, small little videos for, like, every feature on j logs, which is very easy. Speaker 5: We gotta do that. If Speaker 8: you're not making any annotations for Lisa, it's such a low hanging fruit. Just make sure your content is way better, for your end users. Not even just, like, marketing gimmicks. Like, users watch videos much more than they watch, like, the old text. Mhmm. Speaker 8: And some ways that you get ideas. Slide decks, I think people should never make slide decks by hand anymore. You can make great animations, pure libation of PowerPoint slides, and my cloud will just rip through that super easily. All the talks that I got through in as that and, I see other people give now, there's, it's transitioning. People just switch to animated slide decks on that are built not with the slides, obviously. Speaker 8: Just like React. It's a good reason to go build that. Speaker 9: It's only available in the app Speaker 8: right now. Right? On design? Speaker 9: Yeah. You can't call it via API. Once you Speaker 8: once you get it, though, you can just take the stuff and just have it generate code, and then you just use cloud code or codex or whatever code you need you want. Right. Like, the slide deck thing is not a cloud design thing. That's the thing Speaker 0: we were doing before. Was the video's cloud design or is that Speaker 8: We did some snippets through that, but, that was we tried cloud design, but, like, the video side doesn't really make that big of an improvement. It helps, but you can get, the videos are just through, like, asking to make, like, coding based videos. Yep. You can programmatically generate videos, like, feature blogs really trivially. Mhmm. Speaker 8: That's that we're dev tools, so it's only different for other folks. Speaker 0: You're not dealing with crazy animations or graphic design or Speaker 8: much of that? Well, no. It does. I wish I could screencast a video. It's really freaking easy to go make really high-tech videos in under 5 Speaker 2: minutes. Yeah. Speaker 0: We could we could go and check out your service. Speaker 8: I'll send I can send a foundation slide. This is the most recent 1 that we made. Speaker 12: Actually, 1 of Speaker 8: our interns made this, and we didn't even know. Speaker 0: Is it under that video right there? Speaker 2: This 1? No. That 1 is Speaker 8: the old version before we had the. Speaker 0: Is your block out? Speaker 8: There's there's stuff at top point in a couple of them. Let me see. Let me find the ones. Like, this 1, for example. And for the sound side of my. Speaker 9: Put it on the meeting chat. Speaker 0: You just hear it. I like the air dryer. Yeah. Yeah. Find me out of a remote page. Speaker 0: Someone's gonna pick it up. Speaker 6: Just air drop it to everybody. I'll Speaker 0: figure it out. I'll get it. We'll get it posted. Technology. Okay. Speaker 0: So solving air drops is the next thing. Speaker 6: Mode. I've used, remote mode. Speaker 4: I don't know if it's, remote mode. Yeah. Remote mode. It's Speaker 6: it's really good in my opinion because it combines a lot of technologies I'm familiar with, like React, infrastructure, and code, and and and so that Speaker 2: I use a lot. And then it doesn't lock you in really well. Yeah. So you can kinda just perfect. That's for video? Speaker 2: Yeah. Yeah. I just heard about that. So it's been, like Okay. Speaker 0: Do you do you mean by not log in, like, it's open source and you don't have to pay them? Exactly. Speaker 8: I just have to be on Slack. Speaker 9: Oh, shit. Another useful video is Speaker 1: we'll have to use in things field.ai. By any chance, heard of that, but you can make, like, movie theater quality videos, like, $30 in the AI credits. Speaker 2: What's it called? Higgsfield.ai. It lets Speaker 1: you train character locations that's consistent. People are making, like, full on feature films with this thing. And there's their website won't make your eyes bleed, but if Speaker 2: you get past that, I think we can Speaker 6: I just got it? Speaker 1: For a whole room of people. Yeah. Higgsfielg.ai. Speaker 2: But it's Speaker 1: it's insane. Finally, I'll show you the video. I just need, like, Speaker 2: 2, 3 minutes to go over that. Speaker 0: Eric's after that. Speaker 2: Wait. What's the Speaker 0: so this is a a video that I have planned us. Speaker 2: What what tool did you use for this? Speaker 8: So you just use cloth to generate slides, and then you use your replay and, like, I don't know. They I think you use, like, some video programmatic Python where they're, like, be played programmatically Mhmm. In the, into the video format, like, FFmpeg APIs and stuff. Speaker 0: There are people using this for, like, presentations or just, like, I guess, anything. Speaker 8: And there's yeah. There's a lot of artifacts. Just be very clear. Right. Speaker 4: Wait. Where is the artifacts? Speaker 2: You don't see them? Not really. Not in this. Speaker 8: At 1, like, you would never make a video of this many animations. Speaker 2: Yeah. But, like, in the actual code, it's showing up. So there are effects. Right? You you meant visual. Speaker 2: Like, visual artifacts. Speaker 8: Like, see how the a against the syrup, the a against the top 5. Yeah. Speaker 2: Right. But, like, if it's telling you PIP install BAM will p p y. That's correct. Yes. Speaker 8: I mean, you validate the contact. Yeah. Speaker 4: It's getting the important stuff. Right? Speaker 2: Maybe it's Speaker 4: doing a lot of animations to hide the artifacts. Speaker 0: I mean, I didn't notice anything at all, the quality. Very cool. What was the other 1 that you mentioned? See it once you Speaker 2: I'll just see it. Acefield.ai. Speaker 0: So plenty of things to go play around with. I did. Yeah. Speaker 2: Is this being captured? Yep. Speaker 0: Yep. Yeah. And I'll I'll send that out. I'm just not fast enough to put it here, but, I'll send that out after the event as well. Anything else? Speaker 0: Bernie, if you don't mind, your chair is pretty good. 1 more. Let's talk about the bug design. Speaker 2: And then they have skill, Speaker 0: to take that and move over to other tools. So I would expect that to be between the Speaker 2: 3 tools to do that skill. The thought here's how Speaker 0: to make our design dot m Speaker 2: d file work a lot of times. Speaker 0: Yeah. So they they they have a tool to make these things and that has a port of other issue. It's not here Speaker 2: to tell you what LMT is about. Speaker 0: It's not even that hard. It's pretty much for the last Yes. Yeah. Yeah. Yeah. Speaker 0: Yeah. I mean, it's not an arms race. I stopped being a test to Google for a week. They all copy each other now. Yes. Speaker 0: Yes. Do you think Speaker 2: you're good? I would highly recommend clock assignment. If you spend the time to train it on your design system, it's gonna be about 2 hours, but they actually have a really unique UI on training the system on the design, on your design system. They kinda have, like, little step by step from a little bit saying, like, here's your head level. Is this right or is this wrong? Speaker 2: And you can give a feedback. Here are, like, your navigation, you know, tab. Is this right or is this wrong? You keep giving a feedback in the next few hours, but once you get it all set up, first, that becomes your, like, all design system to take this to your other code, basis. But then every page your design after that is pretty close to that design system. Speaker 2: And, we have some cool features. 1 is if you click anywhere on any design element and just comment on it, and then we'll just give feedback just to that design element. And also have a feature called tweaks, which is essentially, like, you turn that on and you can make almost, like, different versions or different states of that design. So, hey. What's it look like for logged out versus logged in? Speaker 2: And you feel like going between. But the coolest part of this whole thing is I'm not a developer. I could I'm more of a designer. You have to see why I'm designing. I could design the actual brand and then give that to my developers, and it's pretty much done. Speaker 2: So, like, we had some of our other employees that are not designers or developers actually give feedback to design. We could actually collaborate on a design with multiple people, and they just comment and things just change. And the idea now is that, like, our entire app, we will be able to actually make changes without developers. The developers will just review this stuff in the end. The front end will be covered a Speaker 0: lot with any of the developers Speaker 2: versus having the handoff between Big Line and developers. I should Speaker 0: be able to work with the Figma. Let's see. Wow. What is your company? Speaker 2: Mark, what's Mark the permission. Once you verify the design system, how do you put your in place? How do you make sure it's consistent? Like, if you like, take care of the design system. But I don't just need this. Speaker 2: You don't need to make sure they have it. I mean, it's just a typical review process. Right? So as I was working with a colleague, she was giving feedback and features, and then I could see all the changes she made. So I can Speaker 0: see your email. I'm what? Kicking off all Speaker 2: of my links. K I mean, kicking off as in, like, She made changes, I made changes, and Speaker 0: then we kind of put a review that Speaker 2: will take time to look. Let's starting next time. We haven't done multiple projects just because it's only been Speaker 0: I don't know. I don't know how you're performing. Speaker 2: We've been working on 1 Speaker 0: project so far. It's been just about Awesome. Speaker 2: Yeah. For years, I think the the the we'd sit down and the designer didn't even do that. I always read because they show you something, and then it's developing, you know, that there's a 100 variables, screen sizes, elements. And they give you 1, and then you just have to go back and build it. Yeah. Speaker 2: Yeah. Yeah. That was 5. So that's the design for screen having to think of all different things. So 1 thing that's awesome, I use Whisper. Speaker 2: So now I just talk to the design. It's not like, can you also add, you know, the empty state? Can you also add the what do you do state? Can you and I'm just talking. By the time I'm done giving 1 note, it's done fixing the others notes, and now I'm, like, reviewing those notes and get them to like, it's a real back and forth. Speaker 2: It's very seamless. So you have to go into cloud design with CMS. And the they're not gonna Speaker 8: take all of that. Speaker 2: I was doing now. It's not part of the need of that. So it's Speaker 6: you know, some first version, so we are a little bit better. Speaker 0: So we need you. Awesome. Final call. 1 more 1 more topic. Anybody? Speaker 0: Yeah. Speaker 5: If you don't have a design system and you're not a designer, there's a group of 3 b's. Yep. You guys have a. Speaker 0: From a say more. From, Speaker 5: It's Twitter for designers. Yep. Speaker 2: How Speaker 0: does this help with the designs? Yeah. Speaker 5: It's ancient. But, if if you need inspiration, you can just search for any type of topic, and, you can see what other people are doing. Speaker 2: Is it still vibrant, or is it, like, Stack Overflow? Everything adds that. You know, it's it's been Speaker 5: the best source so far. I'm sure there are other, folio sites from Speaker 2: Yeah. No. I I'm not I I love Dremel, and I'm on Dremel, but Speaker 4: I'm just curious if it's, like, still vibrant. That's that's normal. Speaker 5: That's yeah. Yeah. I I I can't tell you. So not with Dremel. Speaker 0: And I think there is a foundations company that was actually working on the design review problem as well. I'll look Speaker 5: at everything out there. Speaker 0: We can look afterwards. But, like, the the constant testing that you do to drill a different use cases of your app with your your system. I've used to work in both in the past for React, but, there are they're they're very real process, especially with how much files and code we're pushing. Speaker 5: Alright. The reason why they can release 4 7 design of 4 7 is the first model that they talked about doing visual analysis. And so that means that if you can point it towards what it should be looking for, you can say, hey. Let's look at these pictures, but look for these specific things here. And it actually does a pretty good job at its file finding. Speaker 5: Very cool. So I'm using it for, 3 d reconstruction, and it can actually, look at 3 d models. Speaker 0: 4 7. Ope. It's 4 7. 4. Speaker 5: 0, ope. It's 4 7. Yeah. Speaker 0: That's 1 of the first good things I've heard about more 7. Speaker 2: I'm working on it as a A lot of hate right now. Speaker 5: Security issue that you have to, like, that that was poorly worded in the system prompt. So you have to overcome it by saying no. What I have is not Malibu. Speaker 2: Did you already ask who's using 4 7? Speaker 6: Is that Oh, yeah. That'd be Speaker 0: a great question. Who's using 4 7 daily as a daily driver? Who are we? Speaker 12: There's a half off pro for, like, a week. Speaker 0: Like, that's about me. Okay. Cheaper. Speaker 2: And and then for that, did it just does it just, like, Speaker 4: rank through your credits? Speaker 12: Yes. Or I I Speaker 2: I feel emotionally hurt with that question. Speaker 0: I use it because I can't fall back and call it 0 46. Like, that's Yeah. Pretty unlocked in. Speaker 2: It that's very literal. It has yes. Speaker 9: Very literal. I haven't seen an acceleration of credit use like others have seen, and I haven't not a single instance of it claiming there's some security problem Speaker 2: with what we're working on. I haven't come across any of that. Yeah. You could say I don't know why. Speaker 0: So I'll do that. So I know for 7, supposedly, you put more token because the new token embedding, but I run, like, a follow-up to Simon Williamson, Speaker 8: elegant benchmark. Yeah. Speaker 0: It uses anywhere from, like, 10 to 40% less recently tokens than 4 6 5. So even if the token is more expensive, like, it's using a lot less reasoning tokens. Speaker 9: Well, they they reduced the the cache time out of it from an hour down to 5 minutes. So if you can keep a prompt screen going and get your thought train going within a 5 minute window, it's super efficient. If you don't, if you're beyond that, then the cache times out and it has to reload everything, and then it might use more. Speaker 0: Oh, thanks faster. Or use Westport. Speaker 9: Yeah. Yeah. Exactly. Speaker 0: Alright. Well, thank you, everyone. We've captured that, hopefully, and can get that without, later today or tomorrow. So first, it's up for Thomas, Nicholas. So while you get him Speaker 3: set up, you can get some Speaker 8: So Speaker 0: We hear your awesome music. Speaker 2: Awesome. Okay. Sorry about that. Speaker 1: No. So, I built, an AI interviewing tool that calls the new to these people and then turns that into marketing collateral. The founder of that business, and so I'm gonna talk a little bit about kind of the coding complexities that we've faced with some interesting learnings we've had and gone through the whole process. I started about a well, we'll call it over a Speaker 0: year ago. What's the name Speaker 1: of the company? Oh, InterVidroid. It's called InterVidroid. Started over a year ago. Started to build WebSockets to handle the real time audio for the Jeep, chip and AI, API, integrate with Twilio so you can call your phone and interview directly on your phone. Speaker 1: Got a bunch of early users, and then it just stopped working. It took me, like, 2 months. Basically, gave up on it, but then I have to rebuild the entire back end of it. And it turns out there was, an encoding issue between OpenAI's, API and Twilio at the time. And so if you're considering building real time APIs that are using phones with Twilio, don't use OpenAI and Twilio. Speaker 1: Which is actually I created to 11 labs, so I use a lot of 11 labs for the back end in terms of actual, like, you know, real time intelligence integrated nicely with Twilio. There's a lot of really cool things I could do managing the WebSockets myself in terms of memory and running AI agents side by side. So the reasoning models can analyze the conversation in real time and then injects, you know, the proper proper time without those latest issues. But it wasn't worth the headache and, obviously, the issues with, the a API and Twilio, forced to go down Speaker 2: that path. So the things Speaker 1: about this that I wanna chat about, 1, is, content generation. And so a big part of this product, obviously, is producing content that people actually wanna read. So it produces emails, blog posts, LinkedIn posts, etcetera, all based on that conversation with, you know, whoever's on the other end of that line. To make that content actually good, we do a couple things. But the biggest thing I think is is kind of re reevaluating on itself. Speaker 1: And so, every time we produce content, what we do is we use, Gemini. We use OpenAI GPT, and we use Claude. And we give them about, like, 12 different prompts, writing styles, how to create that content. And so we've produced about 36 pieces of content. And then we have a rubric that we use to evaluate every piece of content. Speaker 1: So for our LinkedIn post, you know, does Speaker 4: it have a good format it correctly? Speaker 1: Does it have the right tone of voice? Is it using words that nobody wants to see? And then once it chooses the top 3 pieces of content, it then generates 1 additional piece using all 3 word feedback. And so, building that kind of real time evaluation loop that it does every single time, we found tremendously improves the overall quality. And, like, you know, this thing's pretty steps to pieces of content, and there's definitely a huge improvement when when we have that on a real time self evaluating loop, using different agents and different parts of that experience. Speaker 1: So that's really helpful. The other big learning, I think, for us has really been, everyone uses databases for everything. This is kind of an anti pattern, but we actually use just blob storage. Things like s 3, Azure blob storage, file storage, databases. You know, you get locked into the schema as a pain in the ass to do migrations, update the schema. Speaker 1: There's lots worth of, you know, really useful things for databases, but they're not always the right tool for the job. And so in this case, whenever you create a wherever they interview, we have a transcript file. We then go and read the transcript file, generate a bunch of new collateral files, and the whole thing runs really quickly, which if you if you end up testing, you'll see. But more importantly than that, like, it's so easy because as you wanna add more metadata, as you wanna change schema, as we wanna do anything like that, you know, just changing text files and CSV files and things like that. So I know that's a huge anti pattern. Speaker 1: Everyone who's going to build an AI tool, the first thing they wanna do is connect it to Speaker 5: a database and try to from Speaker 1: a security perspective, but also just schema and and kind of, you know, wishing you you'd set it up differently in the future. May it makes it as easy as kind of coding locally, I would say. So, those are kind of our 2 biggest learnings, I think, from this tool that that we've had when we when we built it. So that's that's all I got. Awesome. Speaker 1: If Speaker 0: we ask this, if you already have questions. Speaker 9: So is is it is it a real time experience where somebody puts content in and then and then they wait for an output? Speaker 1: Yeah. So the way it works would be, you know, let's say you had, someone at your company who's expert in that industry. Yeah. You could set up an interview with them, and it'll call their phone at the time that you schedule it for. And they'll have a back and forth conversation with the AI agent. Speaker 1: So they'll say something. It'll ask a question about that. It goes back and forth. And then at the end of the conversation, within 60 seconds, it generates all the content collateral, emails you, says, hey. Conversation's over. Speaker 1: Go ahead. So Speaker 9: that's that's the cycle of 36 different versions of what was taken from the conversation or produced then down to, like, 12 agent evaluate blah blah blah down to, like, a minimal set that is then produced from that effect quality. Speaker 1: Exactly. And then share with Speaker 9: Within 60 seconds is what you're saying. Speaker 1: Yeah. Effectively. Because it's all in parallel. Right? So we don't like a lot of, like, the LinkedIn post, like, generating a piece of content, regenerating David, regenerating it, you know, for an output that's maybe 500 words, 300 words, like, the link is very quick on the APIs. Speaker 0: Got it. Speaker 9: Is that a supervisor architecture with agents? Or, like, what's the No. Speaker 1: I I I shouldn't use the word agent. It's not agents. It's a it's a fixed workflow of of kind of Speaker 2: how we process through them. But but Speaker 0: you use for the evaluator of the rubric. Like, if you said there's a rubric that you have Yep. And so, like, there's a bunch of 36 Speaker 2: pieces of content, and then another model use Speaker 0: the rubric. What do you use for that 1? Speaker 1: I think I use, like, OpenAI for that. But we use all 3 models, Gemini, OpenAI, and Claude at different points. We just find it adds more diversity to to the outputs. Speaker 0: I was really curious about your blob storage. I love an anti pattern. Speaker 1: Yeah. It's been awesome. I mean, it cost effective. Yeah. It's basically free. Speaker 1: That's the best part about it. I have, like, 10 different apps that are all, like, useful tools right here and there. And, like, 5 to 10 different databases, I'd be spending at least over $300 a month. Or they're all used to the same database server, which is, you know, it's all dangerous. You know, if 1 gets 1 gets compromised, all my data is compromised. Speaker 1: But Yeah. Literally with blob storage, I don't I think I don't think I David Microsoft a single dollar. So between that and Azure functions Right. Speaker 9: So it's all on Azure. Okay. Yeah. Speaker 1: I yeah. I used to work at Microsoft, so I'm Speaker 0: just asking. Got it. Speaker 2: Got it. Speaker 4: Yeah. Do they have Speaker 9: a blob storage vector solution like AWS as s 3 vector? Speaker 1: You know, I think they just recently relaunched 1. I haven't looked much into it. Speaker 2: So you haven't used it yet? Speaker 1: No. Okay. Speaker 2: So what so you just keep the entire thing is, like, up to all of Speaker 4: the of the entire conversation, and then you can't search through that because it's not vectorized. Right? Speaker 1: Oh, yeah. I use a separate vector database. I can't remember the name of it, but they have, like, a really generous free trial. So Okay. I'm running I'm running this thing on pennies. Speaker 2: Awesome. Alright. Speaker 4: Oh, what what what was your name again? Nicholas. Nicholas. Okay. Speaker 10: And, also, how many times have you got interviewed by your boss? Speaker 1: Probably 50 times or so. Speaker 4: Okay. I Speaker 1: use it a lot for my own LinkedIn content, so I'll be, like, mowing the lawn or something. I'll have your pizza and do an interview or something like that. I hate just sitting down and writing for LinkedIn. Rather than I don't know. Okay. Speaker 1: But wait. Speaker 4: Wait. You don't wanna remind yourself. Speaker 6: You know the graphs No. You talk Speaker 10: the same time, and then Speaker 6: you come back inside and the post is ready? Yeah. Exactly. Okay. Speaker 4: Does it just post it immediately? No. No. Speaker 0: No. No. Speaker 1: I wouldn't recommend that. You do have to. Speaker 2: Deep tank. Speaker 0: So, like, it's pretty tight. It's hard to kind of get a good e dial out. How do you email the business of your entire pipeline? Speaker 1: Can you say that 1 more time? Speaker 8: How do you get out the goodness of Speaker 0: your entire pipeline as you, like, get away with just human review or do you have something in place? Speaker 1: Human review. I think that, you know, I use it enough for myself, and I have other clients that use it as well. And so we're getting constant feedback looped. There's so many pieces of this thing that I'm working on at the same time that, like, when I touch the content side of things, I'll go and evaluate everything, and then just move on and come back months later. Okay. Speaker 2: I think it was smart for you to do a fixed pipeline because agents want to have a lot of time. You can have a, you know, fixed your time and also costs. But, also, you, you know, manage quality. So it is actually pretty smart. Appreciate it. Speaker 2: Yeah. Speaker 1: I mean, I think it's like it's like a very known workflow. So why would I leave it up to an AI agent every time to decide? And that's when BI agents are really great when, like, the workflow is, you know, they don't know exactly what questions it needs to ask. Kinda like that analysis example I mentioned earlier. It's like Speaker 2: they were just going crazy answering all sorts of questions. I don't know. So 1 more question. Speaker 13: Is if you think about your Speaker 2: best customers, is this creating new use cases for things that would have been too expensive and too time consuming to do? Or does this slot into existing traditional use cases like customer discovery or, yeah, podcast type question? Speaker 1: It's It's a little bit of both. And, I'll I'll tell you what I'm focusing on right now. But so there are people who do like, marketing agencies do content interviews all the time with their clients. They do a very good interview over a Zoom call. They record it, take notes. Speaker 1: You know, maybe they throw the transcript into chat g p t, have it write some blog post, and then they get copywriters David and they go post it. But that doesn't work super well because clients always miss those meetings, so those get canceled or rescheduled. And so we're trying to kinda solve that problem, and that's that's where this was born out of because, I do a lot of marketing work as well. And so, like, a lot of my clients all use this, so I'll have to talk to them. Speaker 2: And that's the thing that changes there is the persistence. The persistence can be considered a Speaker 5: huge 1. Right? Because they Speaker 1: can do it, you know, on a 20 minute bar ride. They can do it at night at 09:00 at night or bother home in the hot, whatever. The other side is it opens up now for, you know, like, the attorneys at, law firms, for example. They very possible for marketers to get, you know, them in the same room with them. But this will call the attorney, you know, on their drive home or whatever. Speaker 1: So the attorney can do whatever they want and reduce that cycle. Speaker 0: Well well, we have to we have to go there. Definitely continue the conversation effort. Thanks. What's up, miss? Speaker 7: I'm, I'm Caleb. Speaker 2: I'm part of the 4 team Speaker 13: of Microsoft Labs. I work with the TA, Alex. Here's Dan as well, and Sahana over here. Microsoft Labs were an early stage venture let me do it. Early stage venture fund Speaker 1: in the startup studio down Speaker 7: in Microsoft Square. They're gonna Speaker 13: be pre CNC investing and also we trade our own needs. So what I'm gonna show you today is the experiment we've been running. Last time I was here, actually, I showed you some newsletter that Speaker 7: I was working on, which is Speaker 13: just still images, like, what we need. You know, we kinda did Pedersen, like, what are other businesses that lend themselves to be brought without Speaker 7: a human out there? And 1 of Speaker 13: the most the most lucrative 1 that we found was really AI Speaker 2: oh, it's me. Speaker 13: Alright. Screens. Speaker 2: I'm gonna share Speaker 13: your spot. Cool. Speaker 0: And then Speaker 13: it's open terminal. Was ecommerce brands, you know, in terms of, you know, we look at digital products, courses, things of that nature. So we did it was like, hey. Can we start a brand totally with AI? So the first thing is like, well, what do what do we sell? Speaker 13: So I had Claude and Gemini and Brock all do deep research, you know, look at the market, and I said I gave it some constraints. I said, hey. We have to sell this product through Meta. So we wanna be able to sell it a demographic that has a lot of money, disposable income, and also who we can hyper target through the Andromeda engine. So I gave it just these constraints. Speaker 13: I let it go for a few hours, and it came back with this entire brand brief. It found a domain that was open to before it picked a name, so we have the domain sevend.com. But, basically, this brand, Sevend, is like well, it's it's kinda like this LED face mask, you know, microcurrent device, and, like, I think, like, a cryo roller for women aged about 22 to 35 who live in cities. So, we came up with this whole thing, you know, tone of voice, copy, color scheme, here's the ideal ICP, and, basically, everything you would need, like, you know, how to run meta ads, you would need to, like, run this brand as a human. So from there, I basically fed it into this landing page engine. Speaker 13: I go right here. We can see. And then the second thing, it actually this is logical. We actually it actually went and looked at, like, different drop shipping websites in different sites and, like, found the products on, like, these drop shipping APIs. So they also use that as constraints. Speaker 13: So this is an example. Speaker 1: I found, like, this thing Speaker 13: from China. It's $6.70. It did a bunch of market research on how much this product goes for, and it basically figured out we could sell this for $89. And you guys laugh. Actually, most people sell this for 129, but we don't have humans, so we don't need to pay for labor. Speaker 13: So we're actually giving you a discount if you buy from us. But the first thing it did actually was it generated these on brand, like, lifestyle kind of product images. So, really, I just fed, like, this image into nano banana, not g b t 2 image. And it generated, like, kind of these on brand images. And then also generated, like, this entire, like, well thought outside to kind of fit the brand voice. Speaker 13: So, you know, this is we're able kinda go from just like, hey. I just wanna start an ecommerce brand to, like, alright. Pick the right products that are high margin going after the right demographic to a fully functional ecommerce site. And I think, like, you just talked about, we have this all working on CloudFlare workers and d 1. So we could spin up 1000 stores for no additional cost. Speaker 13: It's only we only pay people to use it. So now we have, like, this product. Next question is, like, how do we market it? They're like, yeah. I think that's why everyone's here. Speaker 13: How do we do go to market? So using Speaker 0: let me show Speaker 13: you right here. We start because we wanna get we have this constraint. No humans in the loop at all. So what we did was it's frozen. Speaker 2: Just does. Speaker 13: Here we go. So what we did was we actually fed in that product image we generated, and we we we generated this video. So I'll show you the video, and that'll tell you how we built it. So this is all AI generated. It actually looks a lot smoother than my laptop. Speaker 13: There's lag. That's that's the Wi Fi, not AI. But Speaker 4: Is there actual voice? Speaker 13: Oh, yeah. There you go. It's plugged in my laptop, but I'll show you later. Yeah. There's there's voice through 11 labs, and we also generated this for a different brand. Speaker 13: We have, like, this this cinematic 1, which I'll show you this brand after. It's another 1 we we came up with. But we're able to generate these super high quality David videos. The way we did this was we start by taking the rep product reference. We then generate, like, a really, really high quality first frame for the video. Speaker 13: We then give it kind of a 3 to 5 second clip of, like, what this should be. We generate 2 to 3 of those kinda we do jump cuts, and then we stitch them all together with FFmpeg, and then we do some smoothing with FFmpeg. So, actually, the most underrated part of this pipeline is, like, the manual video editing. That's what make this possible. So I'll show you this diagram right here. Speaker 13: We can see. So you take the product slug. You write a creative brief. You write a narrative of what this ad should be, and then you start by just generating those first frames. And you kinda you keep it a loop. Speaker 13: So a lot of times, you'll see, like, an extra hand or, like, an extra foot. So what you do is you actually have, like, the VLM look at the image in the first frame, and it just keep projecting until it meets, like, these really strict requirements. Then from there, you basically feed them into the first frame of your David. And then from there, you actually I use the video the Gemini David API to actually watch the video and, like, look for the abnormalities. So, yeah, we have a couple brands. Speaker 13: So we just had we also have, like, this, podcast brand. This is like a GoPro for pets. So it's, again, it's another 1 we started. Again, this is all made by AI, all these images, all this stuff. And we for each of these brands, we have a ton of different, like, personalized landing pages. Speaker 13: So for this 1, this one's, like, for City Parent. So this 1 goes to that ad I just showed you, and we're gonna test 15 different landing pages with 15 different pieces of content to really nail, like, where the ICP is. And the hope is, you know, we can try tons of different products, tons of different creatives, tons of nice to be kinda narrow down, like, finding small hits of what in in these different CPG brands. Speaker 0: So yeah. Speaker 2: Cool. Speaker 4: 2 part question. 1, so it seems like you're using all external models. Like, you're not trying to be clever with hosting yourself. You're not Speaker 13: Absolutely not. Yes. Speaker 4: Okay. Okay. So the second question is, what is it cost to run the AI to create this full fledged drop shipping whatever business? Speaker 13: Yeah. So I was using just cloud code on my laptop to write the code for the website, so I don't have the token cost for that. Each video cost was about 2 or $3. So nothing's going to ballpark Speaker 4: is, like, $5 or less. And Speaker 0: then Yeah. Speaker 7: It's $5, Speaker 13: and there's no infrastructure cost. So it's, like, $5 to launch brand, basically. Speaker 2: Yeah. Speaker 4: Hey. Guys coming for a solid man. Speaker 13: We're cooked. No. Yeah. Speaker 3: Yeah. How much do you Speaker 0: have to spend in ads to, like, to get enough traffic to validate those? Speaker 2: So there's a couple Speaker 13: of things. You can get signal on, like, a $100 in ad buys. So the I did another experiment for a different thing. We ran 7 different creatives. Just, like, you know, I did about a $100 per creative. Speaker 13: And very quickly, you can see the 2 or 3 that win even on day 1 before you spent your budget. So what you do is you just kill the ones at the bottom half and put all that budget towards the top half. So you can get signal on a very little spend on MetaHub, and you can just double down the ones that work. The real key is, like, trying a bunch of creatives. Yeah. Speaker 12: How can you tell that your product can actually deliver what the dropshipper is gonna make? Because the the the quality of drop shippers is all over the place. So you might make a promise that you can't deliver. Speaker 13: So I think so, actually, for this podcast 1, we actually go we have a relationship with the factory in China that manufacture this. So we actually know this manufacturer. The 1 for the drop shipping API was more just, like, seeing how good research was. Speaker 6: Oh, okay. Speaker 13: So we are act we actually do have a relationship with this manufacturer for this 1. Speaker 8: So yeah. Speaker 6: Okay. Got Speaker 0: it. Yep. So are you Speaker 9: actually are you is part of the pipeline generation of the ads, distribution of the ads, managing the what's working, not allowed, you know, all of those things that go Yes. Speaker 13: So Meta has an API. You can just pull the data, like, from each day and, like, hey. How David the ads perform? And then from there, that's how you make decision to kill or or double down. Speaker 9: Right. And the your AI is making those. Speaker 0: Yes. Speaker 2: Yes. Speaker 0: Yeah. Yeah. Speaker 2: So these dropshippers Speaker 13: Yes. Speaker 2: Before, like, people made the margins here because they understood the in the global market. Yes. Now Chinese have really good models come up. Right? Speaker 13: Yes. These are all so reason c, dance, and cling for video. These are Chinese models. Yes. Speaker 2: So what's what's stopping the dropshippers from building? Speaker 9: Well, I mean, I I've I hope they try. Speaker 13: I mean, what's subbing anyway for doing anything? Speaker 9: I mean Yeah. Speaker 13: If you got access to the Internet, you could you know? Yeah. That's yeah. Yeah. Speaker 10: Are you selling the $5 brand launch? Speaker 13: Yeah. Hooking up the Stripe and sending some stuff up there. This is gonna be a library run Pedersen. We'll I'll pour back to you Speaker 7: guys next time about Speaker 13: how we David. So Speaker 4: Yeah. What's your what's your role as the start Speaker 3: of that? Speaker 13: We've not we've not started, like, running these ads yet. We're still in the process of, like, getting some of Speaker 8: the Anything that you've done? Speaker 2: You No. Speaker 13: This is just yeah. We've not none of these are live yet. These yeah. We're still working the manufacturer for this 1. We just have to sign up. Speaker 13: But yeah. Speaker 2: So just to clear the the video, you that was a human in Speaker 0: the loop, did that or not? Automated. Speaker 13: That was totally automated. Speaker 0: That was Speaker 13: totally automated. All the video, I had no input. I just told it, like, hey. Like, here's I just gave I literally just Speaker 9: gave it's like, go look Speaker 13: at the brand brief and, like, our our targets and build make me videos. Yeah. Speaker 9: The the sub and so the CDance model added the subtitles to Speaker 13: No. That was manual through, FFmpeg after the fact. Speaker 9: Gotcha. Yes. Speaker 0: That's what I thought. Speaker 13: That's don't don't try to do sub, subtitles to the models. Speaker 0: That's right. Yeah. Yeah. Yeah. And for me, the ads, do you need the human approval statement there? Speaker 0: Like, don't they the terms require that there's you know, you can't just have the automatic? Speaker 13: No. The Meta API I published multiple campaigns in Meta API. You can you don't need a human belief. Yeah. You you can just if you get get an API token from your Meta account, you can just start running ads. Speaker 13: You don't need you don't need a human to do it. Speaker 1: Yeah. Meta has no ethics or standards. Speaker 0: You do Speaker 13: it to very well, you have to verify who we Speaker 2: are when we send Speaker 7: you a down out, but, yeah, I'm not Speaker 0: down. What? 1 more question. It's on. Speaker 11: What about that? Do you have to bring agents for Yeah. Speaker 13: So, actually, 1 of these I think the failure mode a lot of people have when they try and make AI ads is you just, like, prompt, see, dance. You just prompt the model, and you expect to get a good output. The really thing the the biggest unlock is you wanna generate that first frame with 1 of the image models. Because if you guys have tried the new GPT image model, it's incredible. I mean, truly incredible, like, things it can do. Speaker 13: So you wanna create a really, really crisp first frame. And then there's an image to video model that we wanna use, not the text to video model. That image to video model, you're gonna you're gonna give it, like, the the first frame and then the video, and that's gonna give you much, much better results. And the key is nailing that first frame. So you kinda see this loop right here where it's like, hey. Speaker 13: Extra fingers, orphaned limbs, etcetera. You just keep checking for that, and and you have a visual loop that runs until you get to get hit all hit all those checkpoints. Speaker 0: So yeah. Can you see that on top of it? Speaker 13: Yeah. See, first frame still, visual review, fail. And then when it gets approved, then you send it to by dance for or to see dance for, like, the image to video model. Speaker 7: Sorry. Visual review is you? Or is that No. Speaker 1: That's the l m. Speaker 13: The v it's like it's a visual VLMs are reviewing it. So, yeah, no humans in the loop at all. Yeah. Speaker 0: Last little time? Sorry. No. No. No. Speaker 0: You're good. Speaker 4: For generating the so generate Speaker 2: the first frame, go and generate the video. Speaker 0: Yeah. Speaker 2: Obviously, this cuts of the video. Speaker 0: Yeah. You think it'll improve if you were able Speaker 2: to have several frames for each cut or Speaker 13: So so you we can generate longer ones. I think if you look at we I had I actually had the idea of research on, like, what what are the most engaging types of ads, and the most engaging types Speaker 7: of ads have some kind Speaker 13: of jump cut about every 3 seconds or so. So yeah. But you can you can do longer videos if needed. You just prompt it, and you just give it a longer prompt to what you wanted Speaker 2: to do. Speaker 4: Oh, what I mean is, like, you have the jump cut. So you generate the first frame. Speaker 13: Yes. You can do even yeah. Speaker 12: Series of frames Speaker 1: for Speaker 2: each jump of frame. Speaker 13: Yeah. You just pull the last frame from the last segment, and you feed that as the first frame to the next segment. So you you can you can chain longer frame videos together like that. That's how you see people make movies through Higgs Field. That's what they're doing. Speaker 0: Cool. I do. Thanks, man. Speaker 6: Yeah. Hey, everyone. My name is Jamil Escal. I run a AI first consulting company, and I'm gonna be going over, part of my go to market strategy that's helped me with my last exit and it helped me build, a profitable SaaS product in the video generation space. So I'm actually really glad that Caleb gave his talk prior to me because, my philosophy is actually very different. Speaker 6: I view what was just displayed and, I don't know what Caleb is, but I'll try to take a shot at you. I view it as a race to the bottom in the sense that the reason why the things we do for distribution have gotten so crazy. Like, we're pumping out 50 different AI generated, attractive actor, blah blah blah videos using AI is because we've lost, human relationship building. And I think that's where things are sold at the end of the day. But the way staff have moved is you have, like, a self-service landing page. Speaker 6: The random person signs up. They don't have any relationship with you, and that's why we have to compete so hard on distribution. So I read this blog post, that talked about why SAS is slowly dying, or at least that's the claim it made. And the main thing it said was that what makes SAS so special wasn't actually the software. It was the service. Speaker 6: And we've kind of lost touch with the service. So I I thought, what if you flip the essence in SaaS and we we do service out of software? So I consult with the business owners first. I designed a custom solution for them so they subsidized the R and D, And I know I have product market fit because I'm sitting right Speaker 0: in front of them, Speaker 6: and I know whether they can benefit from this, and then I productize it. And I'll just share a a few of the learnings I've had consulting with, different business owners outside of the technology space. I think we live in a really weird time because AI has made people super motivated to learn about technology, but they didn't learn all the prerequisites. For instance, I've met, like, a 70 year old re real estate agent who's using OpenCloud to send, homes to outbound leads, but he didn't even know what 2 FA was. Like, I I asked him, like, why don't you enable 2 FA? Speaker 6: I couldn't explain to him what 2 FA was. So all that to say, a a lot of people are using AI in a no code way to help benefit their business, but they they don't have any sort of, orchestrator helping them. So, what like, Caleb just demoed was, like, super cool orchestrators slinging together, different AI models to make that website, make those videos. And I found a space to build really powerful orchestrators, to support workflows that people are already using. So all I'm not 1 of her slides, but, this is just showing, like, a a sample workflow that I prioritize. Speaker 6: So I was working with the marketing agency that would film ads on-site, and I was in the age of AI, it's kinda ridiculous that you have to drive to someone's site to film an ad. So I devised a way to, just have them submit a photo of their client and an audio sample of their client and have a finished ad. So I'll skip past all these slides because this is mainly about go to market, but every time you wanna do it with the The TLDR would be how do I do audio without Speaker 0: That's a good question. Speaker 2: I don't know. Okay. Alright. Speaker 6: We have the audio. The the TLDR would be like, okay. You you don't need to film with, the specific person you need in the commercial. You just need, to film any generic person, and have an ad outputted. So guy on the left is the client. Speaker 6: Guy on the right, just a random guy you paid Speaker 0: to act out an ad. Speaker 6: Then I then I prioritized it, and I sell it as SaaS. So this is, like, the the landing page for it. Let me just Speaker 5: log out. Okay. Speaker 6: This is, like, the landing page for it, and it's done pretty well. And I I I think 1 of the reasons why it's done pretty well is because I implemented a few, anti patterns, things I had to unlearn from from working as a software engineer at a SaaS company. I I used to work at Salesforce. A a big thing I learned is when you go hard on the service, observability, capturing metrics doesn't really matter if you can really meet with them once a week and understand their exact pain points. And on top of that, just simple things, can improve their experience quite a bit. Speaker 6: Like white labeling it, like just putting their their company's name.product.com makes them feel like they have a really customized, piece of software. Speaker 8: And how Speaker 6: are we doing on Speaker 0: time? Good. Speaker 6: Okay. And, me yeah. Meeting with them once a week can can reduce the need for observability. And then also learning what I don't need to build by by speaking with them in person is really helpful because if I can see what they can already, do using AI, I don't need to build that into my product. So I was talking with, them, and I learned that they use ElevenLabs, image for background swapping. Speaker 6: And that informed me that, hey. I don't need to build background swapping into my product. So, really, going hard on the service has helped me a lot, and it's helped me especially, not have to do a lot of the crazy distribution stuff that I mentioned at the beginning. Speaker 2: Oh, my god. Speaker 3: Few questions. Speaker 9: So how do you, how are you planning on so you're selling us as a as a service to generate the videos Mhmm. Via a a a website or API or Speaker 6: Via website. Speaker 9: Via website. So people go in, they they compile it on your website, and then extract the video and use it. Speaker 2: Yes. Got it. Okay. Thank you. Speaker 14: Yeah. Who is your mostly real estate agent? Or Speaker 2: 1 more time. Speaker 5: Who who is your customer? Typical customer? Marketing agencies or anyone filming videos, Speaker 6: but they don't wanna visit their client on-site. Speaker 14: So can I just ask the client to send them a photo? I just synthesize all of the videos. Speaker 5: Yeah. So that's exactly the product. Okay. Speaker 11: How are people like, Speaker 7: I have certain mannerisms that I like Speaker 11: to see in my ads when I film them. How are your clients kinda dealing with, like, that's me, but it's not me, uncanny valley? Speaker 6: They just see how that Speaker 0: had to work, and then all of Speaker 6: a sudden, they don't care about that. Speaker 2: Yeah. So you're leaning into service and meetings first first being Pedersen. Have you scaled that? Do you only need to meet with clients once or twice? Or Speaker 0: how do we Speaker 6: I I don't actually feel, a need to scale it because having a consistent client that's doing maybe a 100 videos a week is so much more valuable for me than, probably. I'm not on the Internet and having maybe 5 people visit my site. I like those those 1 to 1 high high ticket customers are just so valuable that, Speaker 2: So you'll see, like, an inventory of the amount of customers, but they're so valuable that that totally works. Yeah. Yeah. Speaker 0: I've kind of worked in, you know, like, 8 years ago. I could see businesses similar to this that were just using teams of people in, like, The Philippines or India to do a lot of the back end work for this type of, like, services business. Mhmm. Are you essentially taking that business model and using kind of these, like, custom AI workflows to kinda minimize the amount of, like, humans that you need to, like, service these customers? Speaker 6: So the only servicing to the customers is, like, that's, initial consultation, understanding their problems Speaker 0: Yeah. Speaker 6: And then meeting with, like, let's say, on a monthly cadence to to make sure they continue using the product Yeah. Fixing any issues they run into. And then the building of it is all outsourced to, like, a slot code or codex. Speaker 0: Yeah. And Speaker 6: then just leveraging, like, off Speaker 0: the shelf stuff. You know what I'm saying? Like, yeah. That, like, you say just be shooting something now. Speaker 6: Yeah. That's great. This is all Yeah. Speaker 0: Cool. So in in terms of managing this, are you so you're not looking to scale it. Who owns the product that gets built at the end of the day? Because from, like, a consulting perspective, typically, you'd sign a master service agreement or something like that, and then the client would own the product. Yeah. Speaker 0: Is that what it is? Speaker 6: So when I work with, like, very large companies, like, I I just sold my last product. I I have to sign 1 of those, and therefore, I can't productize it. But if you're working with, like, smaller companies, that's that's not really applicable. They're not in the business of owning software. I'm not even sure what they do with it. Speaker 0: So so you would own it in those relationships, and then would you scale from a product standpoint, or is that not important because your end goals are different? Speaker 6: So for scaling from a product perspective, I just my blog is I do a really good job for them, and they actually know my other potential customers because all my other potential customers are in their exact industry, which they tend to know. And so organic marketing kicks in really well. And then, like, simple referral programs, like, hey. You get 10% of the revenue for each client you bring in. It scales it, really well. Speaker 6: Sure thing. Speaker 0: Any other questions? Alright. Speaker 6: Alright. Thank you. Speaker 0: David? Speaker 2: David, it's been a few, months since I was, here presenting my, other frameworks that I was working on last time ago. But, ultimately, that was the, inspiration that's gonna work the, tool that I built. Effectively, I'm gonna go over how, you could effectively, adjust your ad copy to to be understood by more users based on their personality type. So and how we're gonna measure that. So, originally, I I built a number of different frameworks over the last couple years. Speaker 2: 1 that was involved in engagement, how to make copy more engaging. Ultimately, a lot of people create really dry, boring copy. I wish I had this tool 20 years ago when I was working with b to b businesses. It's really hard to sell a lot of b to b tools because it's just it's not exact. Then I have a clarity framework, a framing framework. Speaker 2: And the last 1 is personality freshman, which Speaker 5: is 1 of the, you know, Speaker 2: the setup to so, probably the big, the, the personality framework I'm, using is what's called the ocean framework. That's openness, conscientiousness, extroversion, agreeableness, and then neuroticism. It's not it's not MBTI, like a lot of people might be in serious with the, you know, the INTP or whatnot. Those are tend to be a binary, binary types where with the big 5 ocean model, you can be somewhere not introverted or extroverted. You can be somewhere in the middle. Speaker 2: So using that, you can kind of understand certain personality types. So you have your CFOs and finance. They tend to be IC, iconocanciasis, and low oh, I just clicked it on the first word. Then I audicism? Openness. Speaker 2: Yes. Low openness. So, yeah, what what lands risk reduction, audit trail ROI. Product managers are high o, high a, tend to be, you know, more Excuse me. Yes. Speaker 2: You don't need to be projecting? Oh, gosh. What am I how am I not projecting? Speaker 0: You gotta share your screen. Oh, it's a terrible problem. Speaker 2: I guess, listening to that. Sorry, though. My bad. We we go back to yeah. This was the first slide, but, you can, quickly go into the next slide here. Speaker 2: So this one's the slide that I can kind of, review. So, yeah, big 5 buyer clusters by role. Chief financial officer, IC, low o. Product manager, high o, high a. IT, they tend to be high conscientiousness, high neuroticism. Speaker 2: You wanna make sure everything's working. Founders and CEOs are tend to be, you know, high openness and high extra energy. So they have some interesting, ideas. So I built a tool that, helps me, score this. So I'm just gonna start off with just doing a quick, analysis of just a simple phrase, transform your enterprise with AI powered innovation. Speaker 2: So it's gonna take up about 60 seconds. I'm gonna fire off a a follow-up 1 as well. I'm listening to get a you have a message, reduce reporting risk with audit ready analytics. But I have another 1 for backup if, those don't really work out here. So let's see here. Speaker 2: It takes typically almost up to 60 seconds. It pulls down it takes your message, identifies which parts. I have about 33 different, like, all based files, that make up about 1.6 gigs and then 1.6 megs of props rather than put together and sent to the to a quality API to, to do this analysis. It's doing more than just the the, personality type, but, we'll see here. So this, Right on the top engagement for, like, 11:10, 30 platforms. Speaker 2: But down below, we're gonna go look just to the the, ocean type. So here it's saying that, it's high, you know, high on openness, high on neuroticism, and then kind of bid for, for both cocky. It's just the other the other factors. If we switch to the other best gig, we go down to go down here. Let's see here. Speaker 2: Where is it? Oh, gosh. It's, I guess it just cleared out a little bit away. Well, that's, that was exactly how I would wanted to do it. But, conscientious is high, neurotic. Speaker 2: So it changed from, being high openness to high conscientiousness. And, so that's kind of, ultimately, the messaging. But this tool is useful in trying to if you know your target audience, you can take what the top you're using and have it tailor it to the message to the user. I find it really valuable in IDENTIFICATION I go through a lot of different websites, and they tend to be, even big Fortune 500 companies. It's really surprising to see how they're, how they don't target everybody. Speaker 2: And sometimes you wanna target to make sure to target everyone who sees them as well. They've been there. So I don't guess, we probably beat that up with 5 minutes. Any questions? Speaker 4: Has your tool ever gone and scored, like, 10 out of 10 across all categories? Speaker 2: It's really rare to get 10 out of 10 across all categories. Usually, as you start dialing up 1 thing, it kinda starts to add back to, like, different things. So, usually, if you get it really good, it's usually gonna be around 8, 8 out of 10. But sometimes you get that. Speaker 12: Usually, most founders are actually super low and agreeable enough because otherwise, you would talk them out of their idea really easily. Yeah. I'm curious. How do you message people like that? Because they're likely not gonna accept your idea. Speaker 2: Well, yeah. I mean, you would have to. I mean, Speaker 6: I would I would probably throw it Speaker 2: into this tool and ask. Speaker 0: So, Speaker 2: ultimately, it's it's going to, you you can actually put in any sort of What Speaker 12: would it say, actually? They're like, because they're like a 1 out of they're like 1 out of 10 in agreeableness. How many Speaker 2: So they're so they're low agreeable? Speaker 12: Low very low agreeable. Okay. Speaker 2: I as Speaker 0: low. Speaker 12: Yeah. Because Mark Andreessen talks about 1 of the hallmarks of a founder's low neuroticism, low agreeableness. Otherwise, you'd be able to talk about other idea, and, maybe you'd just not do it. Speaker 1: So what's happening behind the scenes right now? Speaker 2: So behind the scenes, it's, so it's taking the the well, right now, it's taking the message, and then it's actually trying to identify what, ultimately, what what is the topic, what is the message. And then it actually goes through my my I have, again, 33 files, but I haven't broken out by sections. Mhmm. And then it's combining the different prompt sections based on what the copy was. So here, vulnerability and our skeptical editor direct, you know, so let's see here. Speaker 2: What repels them? We're all in this together messaging with the best of the more emotion appeals, consensus building. A messaging strategy. Speed with competitive advantage, use direct proof, heavy language, frame of strategic superiority, respect their skepticism. It's and so, yeah, for some of the Speaker 6: skeptics, I think. Yeah. That's great. Speaker 12: I would work on my cofounder name. Speaker 9: So so why the why the ocean is it because the ocean framework is is very flat and thin? Well, the the Speaker 2: the ocean framework's been studied for 50 years, and it has the most scientific backing. While the big thing is with Myers Briggs type indicator and ETI, they don't have the measurement of neuroticism, which has been shown Speaker 9: to change I see. So, like, other personality frameworks like Jennifer Aker and, Jungian type personality. Yeah. Speaker 2: So that's Jungian type because then ETI. Speaker 9: That's a the okay. So so So they Speaker 2: kinda miss out on 1 factor that that isn't in the other 1. Speaker 9: So the the neurata system. Speaker 2: And, originally, when I started building this, I started using MBTI until 1 of my friends was like, what are you doing? What are you using that for? So he he directed me to, like, gotcha. Okay. Speaker 12: Well, it's also interesting because I happen to know all of this too. Is it's also it's very predictive of how someone will respond to a certain message and situation. So it's, like, high conscientious, they're likely to be organized. People low conscientious, more likely to be impulsive and decide on the spot. And so it's like that kind of you can predict behavior based on how they decide Speaker 2: to do stuff. Yeah. So oftentimes, what I end up doing is just I just wanna expand how many people I'm actually reaching. Yeah. Too many people, they just hit 32%. Speaker 2: We added to 50. We have more people, that are interested in product. Speaker 0: Right. We need your mic. Speaker 2: Okay. Speaker 7: I like those songs. Speaker 2: So I feel like Speaker 7: that. What's up, from the last time I showed you how I qualified in inbound leads. I, you wanna know, whether you wanna spend your time on that lead or not. Now, well, let's do outbound. Right? Speaker 7: Let's go and find those leads. And that's what, this is gonna be about. So I'll share a skill that you can, get and use for your own prospect building and, outreach process. But I'll show you how it's, working for me. And 1 thing you'll learn, you know, I have, actually. Speaker 7: So So first of all, since we covered, the, you know, the video generation, I was like, well, let me generate a re motion video for the exact talk I'm giving right now. So there you go. Speaker 4: Oh, wait. Maybe you did it just now. Speaker 7: Yeah. So it's, basically, it's, like, a YouTube I have a skill, again, on my GitHub. It's called Remotion Director. So, basically, you give a a a context, and it just, does the whole thing. It actually does some music, also. Speaker 7: And, basically, it shows that in describe your ICP, and then I'm using the quote management agents to, spin up an agent. And that agent, uses a bunch of that skill that I'm gonna show and a bunch of MCPs to find prospects for this exact event. Right? So I I can see 3 people who actually fit the profile of people interested in this, in this event. So the key components are the manage agents, the skill, which I'm gonna explain, the 3 MCPs that I'm using, and you can get, the skill. Speaker 7: Hit your favorite coding agent to use that skill, make sure we have those MCPs, and, we have an autonomous agent that does is going to have SDR agent. So let me show you and explain how that works. First of all, I'll I'll just, kick off, search. It's, just, basically, a simple UI, to do the process, but we will be looking at what happens under the hood. So, essentially, I just kicked out for that session. Speaker 0: Which Speaker 7: 1 so you can see it, silent here. As you can see, the first thing it does, it actually reads that skill, that playbook, how to find leads for your ICP. Now let's look at that. So, if you know the concept of skills, the the most important part is actually have Speaker 0: this in the Speaker 7: this is basically a safe prompt, safe instruction. So you can see that it basically instructs to load the ICP, parse it, understand what kind of, strategies or strategy to use because I basically I have about 3, 4 different strategies in this skill. Actually, the ones I use internally is, like, 12. But the basic idea is that it tries different search strategies, uses those MCPs. Basically, classifies the complexity of your query, and then, basically, uses the MCPs to find people, through LinkedIn, UXA, generic search, GitHub, yada yada yada yada yada. Speaker 7: And, eventually, when it has a short list of candidates, it, downloads information about them, scores them, and presents you with a shortlist, for you to confirm. As soon as you confirm and validate or provide the feedback, it just goes out and, you know, sources the other 100 or thousand, whatever. You want the end, whatever you have presence for MCP. And, if we go back, to the cloud managed agents, we can see that this is actually gone, live here. It basically, understood the ICP. Speaker 7: Now it basically discovers the MCPs available and the tools that they prefer, the parameters they accept, things like that. Then, basically, it starts, doing the fact. You can see that it decided to use 2 NCPs. The Exact is an interesting 1. Exact is basically is, a new type of search engine, and they have a category of people. Speaker 7: So you can see the exact theory that it does In a couple of seconds, the we can see the Nicolas. There's the tool the tool result. So now it will just process those results, score them compared to your ICP and present you with a with a less similar to what I have here. This is just a free to confirm, and then, we just I instructed to generate more. What's, I think, interesting about this, 2 main first of all, is the that the the the brains are in the scale, and, the data sources are in MC. Speaker 7: That's it. The the key, I'd say, knowledge, competitive advantage kind of knowledge is in the playbooks that you use to find the right people. They're like, for instance, sometimes, say, we have a customer that, is in FinTech, and they are targeting, companies in Europe, with a specific type of AML, like, licenses. So the right strategy is not to search LinkedIn. The right strategy is to go to the European Bank website, download the the list of companies that were assigned that those licenses, and then cross check that information with LinkedIn and, basically, field trip, according to your eyes. Speaker 7: So what I'm saying is that it's I think the whole growth market is, will probably all of us will probably codify the process of going out and finding a prospect, But the the the the value will be in the messaging, like like, we, just, heard. And the other thing is the the the playbook, how you do the search and how we do the initial outreach. And what I realized when outreach for my services for about the sales automation services for about, a month. And, my initial outreach message was along of, like, hey. I'm doing, outreach automation. Speaker 7: Here's, I can find you I can get you 10 companies that fit your ICP, something like that. And the response rate was, like, a half percent. So out of 100, I would barely get the response. But the thing and so I started playing with the, different strategies, and I, figured out that instead of asking if you want to see, a list of companies, I build this agent. So we it basically spans the website of a person I'm going to reach out to, builds the 10 references leads, and sends the spreadsheet. Speaker 7: And what I realized later that I can send the same spreadsheet for the same same type of Speaker 0: company. Yes. Speaker 2: Actually So Speaker 7: it's actually a 70¢. 70¢ is, spent on this kind of, you know, iteration. I can actually send all of the, fractional CFO services companies the same spreadsheet, probably refresh it once a once a month. What's great about it is that, response rate, jumped from half percent to plus 42%. I was like, shit. Speaker 7: So, what I what I realized is that, instead of offering value, just deliver the value. I know it sounds like captain obvious, but that's what I learned. Again, that's how it works. And, I think, please feel free to use the skill. If you please feel free to, improve it, the requests. Speaker 7: And, if you happen to know a better MCPs that, might benefit, that this kind of agent might benefit from, please do get lost in. I really don't know why I can get from a prototype in my flow code that that can service the flow and serving my front end using the cloud managed agents, which they, as Dan mentioned, they announced just this call earlier. So that's basically it. Thanks. Speaker 0: I've heard 2 questions. Is the Speaker 8: name of Speaker 2: the skill, please? Speaker 7: Yeah. So, the name is, the research plugin, and, here's the the, GitHub. My GitHub. And the Remotion director, skill is there as well. So just, this way, review my GitHub. Speaker 7: There are a bunch of different skills that I've shared. Speaker 0: There's also links on the AI Tinkers, meetup site for all the talks, and, those have links in the response. So Speaker 12: Excuse you said you mentioned that your, response rate was 14%. How much increased sales have you gotten down your funnel? Speaker 7: So so it's like, technically, the close rate is about 30%. But when I say a response rate, it it's it means a meeting. So, essentially, it's a It's like a conversion rate. Yeah. It's a, like, a it's a function of that. Speaker 7: The the key problem, I got into is that my schedule, capacity is limited. So now the the that skill I told you about the last time, the lead qualification skill, because it comes in. So now we have a problem on that side of the of the machine. But, basically, I can't yet replace, myself or, my colleagues, in in the b to b sales. But I think for a certain segment of buyers out there, probably, they will actually prefer AI agents than than people. Speaker 7: So we'll see if I can do that, maybe on our next yeah. Speaker 2: Good question. How do you email your skill? Speaker 7: Sure. So, essentially, for the lead, finding, I know what is a good, what what is good, lead. And I have, like, a we call it golden dataset. Basically, a bunch of queries, SVPs, basically, and, types of markers that have to be there. And then we use the as a judge to evaluate those, whether those markers are present or not. Speaker 7: Like, for instance, it could be this company should offer a fractional accounting or for a fractional CFO services. Of course, you could, search by keyword to double check that this week actually does that. But, in my experience, it's much easier to just do an LOM based validation. So we just give a I would say, your goal is to validate, the prospect fund against the ICP. And here's the ICP, and we supply that, from, 0 to 5, 5 being the highest. Speaker 7: And that's that's how we do. And what's great about that this is that since we have the these session walks. Right? And we know, it actually in the debug. You can see the the exact toolbox. Speaker 7: Yeah. You know the step that, when something went wrong. So I used I used, like, a, like, system of hypothesis, of the research to improve the skill based on the validation of the prospect on of the output of the school. Can you see? I I can basically make an, a loop of self improving loop of a skill. Speaker 7: Unfortunately, sometimes it overfits. Like, you know, if on an issue, overfits the skill. Skill becomes too too, fit to specific case. So I need a human in this loop, to actually see look through the skill eventually, but, the whole process works like that. Speaker 0: You're feeding data from, like, the email events, the grade, Speaker 6: through rate, things like that. Speaker 7: I see. In my case, it's LinkedIn. And sometimes, part of my, like, my sales call looks like this. I show a 10 reference leads that I think our agent thinks fits, and I asked the the, estimate to expected estimate to basically rate those and explain me. And if the explanation was on their website, then it means our agent did something wrong. Speaker 7: Like, the the the information was there about the specific targeting criteria. But if it's something new that comes out, human, well, that, helps me understand how to what's what the actual looks like for this for this company. Speaker 1: Does this Speaker 7: make sense? Yeah. Yeah. Speaker 2: Cool. 1 more question. Speaker 11: Yeah. 2 questions. So in the in the skill, do you have, like, a signal trigger library, like, knowing who will be moving up and David the right time or the right kind of, of, I see you going after? And then secondly, how do you also have sequence of outbound messages in that skill? Do you have all that factor Speaker 7: in as well? Let's see if you can also. So, the signals are there, but not the messaging. For the messaging, we have a different skill. But, essentially, this is, focused on the prospect building. Speaker 7: So, if you then look into this, SkillMD, you will see that, when the ICPs Mhmm. Apply, the signals that highlight something undefined. Like, for instance, for this for the prospects for this event, I know that 1 of the signal was that whether this person attended a previous AI tinkers, which makes sense. Right? They are GTM engineer, and they previously attended. Speaker 7: Probably they know they've got the the newsletter, but, I think the the basic idea works. So what I'm saying is that for start up's friends is that raise funding, CFO fractional CFOs typically offer their services to you. Recent, running around is a great signal, and it the the skill instructs to look for that based on the IC. What about the sequencing? So but the sequencing, it's not part of this skill. Speaker 7: Overall, messaging is not part of the skill. It's a different skill. I don't have it yet open source, but probably, Voci. Speaker 0: Thank you. Speaker 15: Terminal terminal message of flow on and in a terminal 2 days each via live singing messages while we speak to different conversations, no network delay. Dan's got the flow in a terminal way. Team is building now. He's a I'm I'm talking to the company where apes are found in. And then your console began to be so clean. Speaker 15: Real time presence updates on your terminal 3 on files, make the navigations with the black binary bundle. That's it. But get concession through the messages. Progress across the globe. GTM and networking with that terminal flow. Speaker 0: Perfect. You gotta thank Brian for all that music. AI generated music. People were talking to class. Speaker 3: We got we don't We Speaker 2: don't build job market schooling, and Speaker 0: I build stupid novel. Yeah. It it makes it so much better. I look forward to it every time now. I think you're on on the hook for that. Speaker 0: So so a 2 years ago, I started building a, LinkedIn agent that would just manage my little kid for me. And, it was hard because LinkedIn is notoriously locked down. But today, we have Claude, and Claude is really good at reverse engineering, source code or code that's compiled and obfuscated and all sorts of stuff. So in 24 hours, I I let it run overnight, and it spent way too many tokens because I didn't have Linus in place. But, I was able to essentially re reverse engineer, what took me a $100,000, 2 years before and a lot of development hours with, Eastern European developers trying to achieve, who were great at this stuff. Speaker 0: And it's now possible, to just do that, super simple and super easy. So, I decided to take that and build towards this future that I would personally wanna see. I've always said that, I cannot wait to be wealthy enough to be off of the LinkedIn. So my, my goal with this, though, is that all the current providers out there it was a a pretty great company. We had, about a 100 really, really happy customers. Speaker 0: And I went out to fundraise for it, and everyone was like, you're crazy trying to take out Microsoft. It's not a VC business that we wanna get behind. So chartered that idea, and I still wanted to help those people as well as help myself. And so I'm now open sourcing, this. And what this is is it's a, command line interface to essentially manage your LinkedIn inbox for you. Speaker 0: So, everyone here probably got a LinkedIn message for me. That was not for me. That was from Claude. Mhmm. My goal and where I wanna see this is, I mean, I've I've made a lot of money off of relationships off of LinkedIn and providing value to people, and I want to scale that. Speaker 0: But I wanna do it on both sides of that equation. I don't want it receiving all of the spam and all these connection requests and spending all this time on that social media platform. I just want agents to talk to agents and figure it out for us, like, kinda like what Hollywood does. It's you people who are talking to my people. And so I think that's coming, with open flow whether we wanted to or not. Speaker 0: But 1 thing that's missing is there's a ton of lock channels that are not open, and you have to go through all these shady gatekeepers instead of just knowing it yourself. And so all man, is a a call out to the creator of Sendmail to essentially just give agents a command line interface tool instead of a skill or anything else, to essentially manage your inbox for you. So I built 2 different things in particular. 1 was I built, the command line interface, which is just a tool that you can essentially, query directly. And what it does is it builds a local GitHub repository or git repository of your entire LinkedIn inbox messages and everything stored locally for you indexed and keyed, signaling, and all sorts of fancy stuff that makes this work. Speaker 0: The higher level work of managing an inbox a lot easier. And then just as a proof of concept to see if it worked, I went ahead and, had clawed by, TUI to wrap that and, added a whole bunch of things into that. So instead of demoing the CLI, which will also be part of the release, next week, The, TUI here, which I thought was fun. I've never built a TUI before. You'll see a lot of the messages, that I sent out directly in here. Speaker 0: So for example, I have an entire command line interface in this TLI or, TUI where I'm able to, like, start new conversations, based on people's slugs on LinkedIn, able to keep up sync. This is live synced as well, so it's holding open a WebSocket, that LinkedIn sends events and messages to. And so it can respond in real time, if I turn that on and build that. Can handle reactions. It can handle all the different types of message content. Speaker 0: It's it essentially reverse engineered LinkedIn's private API that they use for their web product. So and this is all directly from this machine or wherever you put it. You can proxy it or do whatever you want. But the important part was I didn't want that to go through some other providers. I wanted that to be something that I have. Speaker 2: And this is their messaging product? Speaker 0: This is messaging Yes. Specifically. There's a bunch of other stuff that you can do, scraping and pulling data, and that's their mode and what they care about and want to protect. And so I'm hoping that this is a feature that they will accept eventually as soon as they actually build the tool to accept this feature. So, let's see. Speaker 0: I want to try something. So AI Thinkers also has an interface here, where I can pull everyone. So, I Speaker 2: I just create nice time messages. Events and have it, Speaker 0: Sunday, it comes Speaker 1: Then please let you're don't be banned on on LinkedIn? Speaker 0: So I've been I've been automating LinkedIn for since it's sounding, basically. Yeah. I have never been banned. Speaker 2: Oh, my god. Speaker 0: The 1 thing that I would absolutely recommend you do this, I've I've never received a warning either, which is pretty impressive. I've, I've had other people receive warnings that I'm getting paid. But I I Speaker 2: So so be careful. Yeah. Speaker 0: The 1 thing I would absolutely recommend is become a paying customer there. Yeah. So I pay I think it's a $100 a month, the sales nav, and it's a totally different set of rules. They have totally different APIs, but each of the APIs has its own internal rate limits and different things. This even has rate limiting built in, which I'm hoping it actually does. Speaker 0: So, it's great. 1 second. So thanks for coming. Speaker 3: So, Speaker 0: yeah, while that's running, let's see. Okay. It's coming in. We we have reachable. No. Speaker 2: Yeah. Just draw You're sending connection request? Speaker 0: I did not build connection request into this touch 1. But Is that Speaker 7: the connection added or, also to you know, you're not connected here? Speaker 0: So LinkedIn has about 12 different rules of who you can actually message based on different factors, and so this will probably just fail for the ones that it won't work. Speaker 7: I'm sure. I think this Speaker 0: How do I I don't think there's a login committee. That is a whole another, area because of caches and and other anti bot protections that they have. My theory is, and I I do this, so maybe it has to do with never getting a warning. I have a, browser use agent who play LinkedIn games for me, and, there's some mouse movement stops. But I think, like, the original CAPTCHA, it was never about solving CAPTCHA. Speaker 0: It was about how you sound you saw CAPTCHA. So that's my theory with LinkedIn games. Maybe that has something to do with it. Speaker 2: So you made a possible play game so you can automate work. Awesome. Yeah. Speaker 0: If you play those games, you're not missing out. Speaker 2: So Yeah. They're not ready. Speaker 0: Not fun games. I have better games to play. Speaker 7: Does it does it say InMail or just the regular messages? Speaker 0: These are just straight up messages. Yeah. InMail is dead. Yeah. It looks like they're not right. Speaker 0: Yeah. Speaker 2: It was like So the the move here is go add 10,000 people and then set up some automations? Speaker 0: I don't think so. Number 1, you can't add 10,000 people. You can do what is it? 200 a week with some good sales now, give or take. I think the move is to be very, very targeted. Speaker 0: But what I want is this like, the agent that I'm building next off of this for me personally will be able to respond and classify just like it can manage my email at the same time. So why can't you put that on LinkedIn? Like, I think it would make LinkedIn so much more valuable because it's an authenticated source. And, like, they already have, layers of verification stuff. So this is interesting. Speaker 2: I decided to build Speaker 0: a script, before I had it. It it was just using the tool. This must be a 4 7 thing. But, cool. Grow script, and that's doing some PSP. Speaker 2: Not what I would do, but sure. Speaker 1: I think it for me, part of it is Speaker 2: don't you see this? I just really fixated on this idea that you just don't have to go to LinkedIn. You don't have to go to this other service, this other service. Some degree, you're bringing that conversation into a place and a time and a moment, you know, on your low Speaker 0: that I own That you own. That I own control. That's the biggest that's my legal defense for releasing this on GitHub, which is done by Microsoft. So, it's it's I'm not running this on behalf of anybody else. It's their own data. Speaker 0: GPR in Europe, this is perfectly legal to do. They yeah. Just make it hard. I don't think that this is, where where I think the reckoning comes through LinkedIn is their mode, or their business model is eyeballs and ads. It's still, like, 80% of their revenue last time I checked. Speaker 0: And so, I think there is a pay to play or pay to increase. But the the more that we have this out there, yes, there will be, bad actors that try and take advantage of it. But I think with the right levels of intelligence and the right use cases, especially on the receiving end of that as well, like, the market will work itself out, and it will be much higher. Speaker 2: Think about just the, using, Google Calendar and Google Mail with, Speaker 0: OpenBook. Speaker 2: And I don't have to go there, and I get it summarized in my WhatsApp. And all of a sudden, I'm just like, wow. Mhmm. I haven't touched my inbox in a month. Mhmm. Speaker 3: Are you just making this with OpenBot? Speaker 0: Yeah. I mean, this is a CLI tool that you can just on to open call and have it run there. So while it's running okay. There was a few that were unreachable that has gotta accept my bots connection request. Okay. Speaker 0: Okay. So while it's, pardon me. So but Speaker 9: it comes to us you. This is your personal account. Right? Yes. Yeah. Speaker 0: So not a business account. This is yeah. Which is why I think it works in Speaker 6: a yeah. Speaker 0: But we're gonna get views Speaker 2: if we get out there. People start spamming everybody. People already spent. Speaker 0: Yeah. Someone said, does this make LinkedIn less authentic? And I was like, was LinkedIn ever authentic? Yeah. That's true. Speaker 0: So Speaker 2: it becomes like Reddit, so you just ignore that. So so you have to build a tool to ignore that. Speaker 0: I actually am a huge Reddit user because I'm able to very much control a lot of those things. And on LinkedIn, I'm not. So, if you sign up today, you'll get on the, quote, unquote, wait list, and it just gets an email. It's just to get a repose that I'm working on some file touches. I'll actually give you access to it tomorrow so that That's it. Speaker 6: You guys see my Speaker 0: allman.sh. Speaker 2: Good. Automation's letting me know if you on messages you received. Speaker 0: See if it's spam deleted or Yeah. Or it's classified or review it. Or who is this person? Why would I care about this person? Like, I have a mental model for every connection request that I received to be like, oh, this is legit. Speaker 0: Are they like, is this a connection that's worth having? I don't want a customer at the, previous company who used, like, a social ranking, mechanism of, like, where are they in the company? What's the size of their company? Like, this is pretty high level stuff because he had already reached a 30,000 connection, but then on LinkedIn. And so, that was for inbound because he received he's a public issue figure, so he received it fine. Speaker 11: Oh, so I got your message on Yeah. Speaker 0: Is this 1? Yeah. Check your LinkedIn. Make sure. Speaker 11: Because I I can tell this is not from you. Speaker 0: Okay. Well, I didn't even ask the message. I'll be good. But I think that's the layer that comes up this. Like, I just wanted this morning, Yeah. Speaker 0: But that's definitely the key. Speaker 3: Yeah. Question. I was wondering why you don't post Speaker 2: the interface separately or or, like, create an interface for the cloud code and, like, you need to interact with and, like, create a signal or and then as well. Speaker 0: It seems like it's running these sequential Speaker 2: right here. I'm gonna be batching. Speaker 0: Oh, that's just that's just what we decided to do. Yeah. The CLI tool actually has rate limits inside of it itself. So, it's still going. But for example, just grab it out here. Speaker 0: Yeah. Probably won't actually do it. Yeah. Yeah. Great. Speaker 0: Let me still pin. So you don't want to because of LinkedIn's restrictions. You don't want to probably patch it that way. But yeah. So right now, Claude go wrote a script to use the CLI, and I just have this over here. Speaker 0: And we could actually see the messages come through. Speaker 4: So you're just doing this yourself. So if you just take bearer token, a lot, everything like that, or Speaker 2: are you using a file and get them? Speaker 0: No. So this is all built from scratch, reverse engineer it off of their internal API. So that people just put in Speaker 4: their branch into t the t y? Speaker 0: No. So they, you can just type, what's it called? All name and log in. Speaker 2: And then it opens up? Yeah. Speaker 4: It should. It goes back? Speaker 0: Oh, it's open. Nice. Beautiful. That's why I need beta testers. Yeah. Speaker 0: So right now that opens up a browser for you to log in, but that's actually pretty easy to log in. And then Speaker 2: you get the l I a l a t token system. So And it's a lot more complicated Speaker 0: than that because they actually have tons of different, security measures in place now. And so it's not just storing the 1 token when we're using it. Like, it's the. It's actually keeping track of cookies as it continues the session. Speaker 2: Okay. Obvious problems with various API stability. How stable is the API? Speaker 0: So not at all. They have they have tons of shards, so I I'm fully expecting that not everyone in this room will work from this depending on, I've seen it based on where your account's created. I've seen it based on a whole bunch of other factors. So, that's 1 piece that I fully expect, to build into this. The charts are running the API code, basically. Speaker 0: So they all file with a hash, on their API because they have a chartered API as well. And so it's, way too overly complex. And you'll see posts on LinkedIn, LinkedIn engineers complaining about it, which is kinda funny. But, yeah. So the the point, though, is that this system should be able to sell because it did the reverse engineering stuff. Speaker 0: And as it gets better at that, whatever the product can do, the tool itself can reverse engineer your field for that, and it just changes how we think about products that we're building. Speaker 2: How did you Pedersen engineer the APIs? Speaker 0: Sorry? Speaker 2: How did you reverse engineer the APIs? I, Speaker 0: had to go to linkedin.com and a logged in session, which downloads the entire JavaScript file obfuscated file, and I said go figure this out. Speaker 2: Wow. So as you know, a lot of people have different jobs.