UNDERWRITE

A due-diligence analyst you can leave running overnight.

Give it a list of startups. It plans the research, pulls company data from Crustdata, reads the web, and writes an investment memo where every fact links to the exact sentence it came from. An independent checker throws out anything the sources don't support, and disagreements between sources are flagged in red. A Coach agent reads the agent's own neatlogs traces and rewrites its prompts — and keeps a change only if the eval score improves.

HOW IT WORKS

1. Plan — turns a company into 8–10 diligence questions with search queries.
2. Research — Crustdata company record + web search + page fetch, within a fixed budget.
3. Extract — atomic claims, each pinned to one source and a verbatim quote.
4. Verify — quote must exist in the source; an independent judge checks support; values are cross-checked.
5. Fill gaps — unanswered questions trigger a second research round.
6. Write — a cited memo and evaluation form, using verified claims only.
7. Recover — retries, model + data-source fallbacks, JSON repair, checkpoint/resume after a crash.
8. Improve — the Coach reads neatlogs traces, detections and triage fixes over MCP, rewrites prompts, re-runs the eval.

SCOREBOARD

Fact accuracy/precision: agent's fact sheet vs. hand-checked ground truth (founding year, HQ, founders, CEO). The Coach only sees failures from the train split; test is held out. *List-price equivalent; the runs themselves used free tiers.