AIVC Research Copilot
Turns raw VC meeting notes into evidence-checked deal memos in under two minutes.
VC partners leave meetings with messy notes and voice memos; the fund's memo discipline dies in the transcription. AIVC closes that gap — structured, evidence-checked deal memos from raw input, fast.
Paste notes or upload voice; the pipeline extracts claims, runs parallel web-evidence searches on each, detects contradictions and gaps, labels every statement — Fact, Inference, Speculation, or Unknown — and generates the memo. A ground-truth benchmark suite evaluates the pipeline end to end.
Architecture
Claim extraction runs on a fast, cheap model (Claude Haiku); memo generation on a stronger one (Claude Sonnet); Whisper handles voice; Exa powers four parallel evidence queries per memo. Confidence labels come from deterministic rules, not model vibes — the same input always classifies the same way.
Production posture
The pipeline is built to be measured, not just demoed: a ground-truth benchmark suite scores extraction and classification against hand-labeled examples, so a prompt or model change is a regression test run, not a vibe check. Model tiering keeps cost per memo predictable (cheap model for extraction, strong model only for generation), an intake gate rejects inputs the pipeline can't do honest work on, and every claim in the output carries its evidence trail — the memo shows what was checked, not just what was concluded.
The live demo is gated; email me for the access code.