Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
veritract - LLM structured information extraction and verification
veritract is a run-locally LLM information extraction and verification package.
Most extraction tools stop at valid JSON. veritract’s differentiator is the grounding layer: every extracted value is traced back to a character span in the source, and values that can’t be verified are quarantined rather than silently returned. You get typed provenance on every field (direct / paraphrased / inferred / quarantined) — so downstream code can treat high-confidence and uncertain fields differently instead of accepting or rejecting the whole result.
Benchmarked on EBM-NLP 2.0 (155 expert-annotated clinical abstracts, head-to-head vs LangExtract, gemma4:e4b): 87.7% field accuracy, quarantine precision 100% — every field flagged as unverifiable was genuinely wrong.