Finance
Audit-grade company diligence from one prompt
Name a target company and get back a verdict-led diligence memo: financial health, leadership, legal exposure, competitive position, and risks, every claim carrying citations and a confidence grade. A CrewAI crew reviews the research and flags the claims that rest on weak evidence. The app then acts: a PDF memo lands in the deal team's inbox, follow-up questions go back to the live agent, and standing instructions from analysts are written into the agent itself, permanently.
Quick Start
Inputs
- Target company A plain-language diligence request entered at run time (for example, "Due diligence on Perplexity AI for a strategic investment").
- Follow-up questions (optional) Free-form questions after the memo, answered by the live agent with the memo's research context, not from stored data.
- Standing instructions (optional) Rules the analyst should always apply (for example, "always assess AI copyright litigation exposure"), written into the agent permanently.
Outputs
- A verdict-led diligence memo with six investigation areas
- A cited claim for every field, graded by confidence
- An evidence-gap list of weak load-bearing claims
- A branded PDF memo, emailed to the deal team
- Live-researched answers to follow-up questions
- An auditable log of what the analyst has learned
How it works
A 6-phase pipeline.
- Research The Diligence Analyst agent investigates the target at maximum effort across SEC EDGAR, Crunchbase, PitchBook, LinkedIn, court records, and the press: one async run, minutes of real multi-source research.
- Verdict The memo returns schema-conforming with a judgment up front: proceed, proceed with conditions, caution, or do not proceed, with a rationale tied to specific findings.
- Review A CrewAI crew audits the research: acceptance-criteria QA, then a Risk Officer cross-examines the trust claims and lists the load-bearing ones resting on low confidence or a single source.
- Act The app renders a PDF memo with a verdict badge, risk table, evidence-gap appendix, and full citation appendix, and emails it to the deal team automatically.
- Ask anything Follow-up questions re-enter the live agent with the memo's context via multi-turn. The agent researches the answer; it does not summarize stored data.
- Learn Standing instructions are written into the agent's expertise and logged with before and after diffs; every future memo honors them.
Stack
Nimble primitives plus the full runtime stack.
Nimble APIs
What it does
- Web Search Agents One max-effort research agent built from the due-diligence gallery template: hundreds of sources synthesized into a verdict-led memo. Multi-turn runs answer follow-ups; agent updates make it self-learning.
- Trust framework Claim-level citations, confidence grades, and per-field provenance on every result, stored per claim in Supabase.
3rd Party Tools
Role
- crewai Review crew: a QA analyst checks acceptance criteria, a Risk Officer flags weak evidence, an Editor writes the partner narrative.
- claude-sonnet-4-6 Drives the three crew agents.
- supabase Postgres storage for memos, per-claim trust rows, follow-ups, and the analyst's learning log. Free tier.
- fpdf2 + resend The action layer: PDF memo generation and automated email delivery to the deal team.
- streamlit UI: memo view with confidence badges, trust panel, follow-up questions, and the teach-the-analyst page.
- python 3.10+ Agent setup, run lifecycle, and the research, review, act pipeline.
Reach out if you have any questions.
Talk to an Expert