UnderWriter

Autonomous VC Diligence, in 60 Seconds

A Cursor-driven agent that fans out six diligence desks, streams findings to a Next.js UI in real time, synthesises an IC-grade verdict, and queues a wire — or holds it and proposes a mandate amendment as a pull request.

The pitch in one paragraph

A GP types: "Wire $2M to Acme Robotics for their Series A. Lead is Sequoia. 50% pro-rata of our $4M allocation." UnderWriter parses the prompt and SPA, fans six concurrent diligence desks out across Specter, Companies House, OpenSanctions, WHOIS and the fund's own MANDATE.md, streams every citation to the UI as it lands, and either queues the wire or fires a BLOCK — catching, for example, a one-letter typo (acrne.co vs acme.co) on a forged wire-instruction email registered six days ago with a failing DKIM signature. When the verdict is HOLD, a Cursor SDK Composer-2 agent drafts a pull request that amends the mandate so the same pattern can never slip through again. The fund's playbook compounds in git.

The demo catches a $2M Business Email Compromise in 60 seconds that a junior associate would have wired.

The Specter integration follows backend/docs/SPECTER_FLOW.md end-to-end — POST /companies → GET /companies/{id}/people → GET /people/{id} → GET /companies/{id}/similar, the §3b derived fields (prior_exits, stealth_history, departed_subject_company), and the §4 cross-step underwriting flags (founder_departed_before_close, ex_founder_now_at_investor, funding_outlier_high|low). The canonical worked example — Dex (meetdex.ai) closing a $5.3M Seed led by a16z with a co-founder who exited 26 days before close and now scouts for the lead — ships as a third demo seed (dex-meetdex).

Why this matters

Problem	UnderWriter's answer
BEC scams cost the industry $2.7B/year (FBI IC3 2023)	A dedicated Wire Safety desk runs WHOIS + Levenshtein + DKIM + sanctions screening on every wire
Diligence is a 2-week analyst slog of 30 browser tabs	Six desks in parallel, joined at a single verdict, in under 30 seconds
Every fund has a mandate; nobody reads it	The mandate is the agent's source code — pure rule evaluation, deterministic, auditable
Policy drifts; institutional knowledge evaporates	Every override drafts a PR against `MANDATE.md` with rationale and run reference

Technology stack

Layer	Technology	Why
Agent runtime	Cursor SDK · Composer 2	Sandboxed cloud VMs, parallel subagents, MCP tooling, file-editing for amendment PRs
	OpenAI GPT-5.5	Structured prompt parsing, SPA/wire extraction, memo editorial summary (Zod-validated)
	MCP (Model Context Protocol)	Specter, Companies House, OpenSanctions, WHOIS exposed as tools
Backend	Next.js 16 Route Handlers	Single runtime for HTTP + SSE; zero ops surface area
	React 19 · TypeScript 5	End-to-end type safety from `RunEvent` at the seam
	Server-Sent Events (SSE)	Live streaming of `desk.start` / `desk.citation` / `desk.resolved` / `verdict`
Data sources	Specter (companies, people, transactions, interest signals)	The unique data — who's actually leading the round in the last 60 days
	Companies House · OpenSanctions · WHOIS	Registry truth, sanctions/PEP screening, domain provenance
	PDF / EML parsing (`pdf-parse`, `mailauth`)	SPA extraction; SPF / DKIM / DMARC verification on inbound wire emails
Frontend	Next.js 16 App Router · React 19	Mandate / Run / Memo screens; SSE consumer renders tiles in real time
Frontend	CSS variables + dark/light theme	Fixed-width memo template that looks like a real fund document
Validation	Zod	Every LLM output is schema-checked before it touches the verdict layer
Mandate	`gray-matter` (YAML frontmatter) + Markdown	The policy file is the source of truth — readable by LPs, executable by the agent
Resilience	In-memory cache + fixture fallback per source	`DEMO_FORCE_FIXTURES=true` = on-stage panic button
Mock rails	In-process wire ledger	No real money moves; every run is a queued / held entry

System architecture

flowchart TB
    subgraph User["GP — Browser"]
        UI["Next.js 16 UI<br/>Mandate · Run · Memo"]
    end

    subgraph Backend["Next.js 16 Route Handlers · port 3001"]
        RUN["POST /api/run<br/>SSE stream"]
        MEMO["GET /api/memo/:runId"]
        AMEND["POST /api/amend"]
        ORCH(["Orchestrator<br/>agents/orchestrator.ts"])
        SYN(["Synthesise<br/>findings → verdict"])
        MEMOGEN(["Memo generator"])
        LEDGER[("Mock wire ledger")]
    end

    subgraph Desks["Six diligence desks (parallel)"]
        D1["01 Company"]
        D2["02 Founder"]
        D3["03 Lead investor"]
        D4["04 Round dynamics"]
        D5["05 Mandate"]
        D6["06 Wire safety"]
    end

    subgraph Sources["Data sources (each w/ fixture fallback)"]
        SPEC["Specter MCP"]
        CH["Companies House"]
        OS["OpenSanctions"]
        WH["WHOIS / DNS / DKIM"]
        MD[("MANDATE.md<br/>YAML + prose")]
        OAI["OpenAI GPT-5.5"]
    end

    subgraph Cursor["Cursor SDK · Composer 2"]
        AMENDER["Amendment drafter"]
        PR[("GitHub PR<br/>against MANDATE.md")]
    end

    UI -- "POST /api/run" --> RUN
    RUN --> ORCH
    ORCH --> D1 & D2 & D3 & D4 & D5 & D6
    D1 -. "tools" .-> SPEC
    D1 -. "tools" .-> CH
    D2 -. "tools" .-> SPEC
    D2 -. "tools" .-> OS
    D3 -. "tools" .-> SPEC
    D4 -. "tools" .-> SPEC
    D4 -. "GPT-5.5 SPA parse" .-> OAI
    D5 -. "rules" .-> MD
    D6 -. "tools" .-> WH
    D6 -. "tools" .-> OS
    D6 -. "tools" .-> CH

    D1 & D2 & D3 & D4 & D5 & D6 -- "DeskFinding" --> SYN
    SYN --> MEMOGEN
    MEMOGEN -. "editorial summary" .-> OAI
    SYN -- "PROCEED" --> LEDGER
    SYN -- "HOLD" --> AMEND

    AMEND --> AMENDER
    AMENDER --> PR

    ORCH -. "RunEvent SSE" .-> UI
    UI -- "GET /api/memo/:runId" --> MEMO
    UI -- "POST /api/amend" --> AMEND

    classDef tile fill:#0b0b0b,stroke:#666,stroke-width:1px,color:#eee
    class D1,D2,D3,D4,D5,D6 tile

The diligence run, end to end

sequenceDiagram
    autonumber
    participant GP as GP
    participant UI as Next.js UI
    participant API as /api/run
    participant Orch as Orchestrator
    participant Desks as 6× Desks (parallel)
    participant Ext as Specter / CH / OS / WHOIS
    participant LLM as OpenAI GPT-5.5
    participant Mem as Memo store

    GP->>UI: prompt + SPA + wire instructions
    UI->>API: POST /api/run (RunRequest)
    API-->>UI: 200 OK · text/event-stream
    API->>Orch: runOrchestrator(req, send)
    Orch->>LLM: parsePrompt(prompt) → ParsedDeal
    Orch->>Orch: loadMandate() (gray-matter)
    Orch-->>UI: run.init { runId, mandateVersion }

    par Six desks fan out (Promise.allSettled)
        Orch->>Desks: company
        Desks->>Ext: Specter + Companies House
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    and
        Orch->>Desks: founder
        Desks->>Ext: Specter + OpenSanctions PEP
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    and
        Orch->>Desks: investor
        Desks->>Ext: Specter Interest Signals
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    and
        Orch->>Desks: round
        Desks->>LLM: parseSPA() (Zod-validated)
        Desks->>Ext: Specter Transactions (comparables)
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    and
        Orch->>Desks: mandate (pure code, no LLM)
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    and
        Orch->>Desks: wire-safety
        Desks->>Ext: WHOIS + DKIM + Levenshtein + sanctions
        Desks-->>UI: desk.start · desk.citation… · desk.resolved
    end

    Orch->>Orch: synthesise(findings) → Verdict
    Orch-->>UI: verdict { action, confidence, blockingDesk? }
    Orch->>LLM: editorial summary (memo lede)
    Orch->>Mem: saveMemo(runId, MemoData)
    Orch-->>UI: memo.ready { memoId }

    alt verdict.action = proceed
        UI->>GP: green PROCEED bar · "Generate IC Memo"
    else verdict.action = hold
        UI->>GP: red BLOCK modal with reason + clause
        GP->>UI: click "Override & amend"
        UI->>API: POST /api/amend (OverrideContext)
        API->>LLM: Composer 2 drafts MANDATE.md diff
        API-->>UI: AmendmentDraft { branch, diff, prTitle, prBody }
        UI->>GP: PR preview + rationale
    end

The six desks at a glance

flowchart LR
    subgraph Inputs
        P["Prompt"]
        SPA["SPA PDF"]
        WI["Wire .pdf / .eml"]
    end

    subgraph D["Six desks · concurrent · 30s timeout each"]
        D1["<b>01 Company</b><br/>Entity exists,<br/>active, on-script<br/><i>Specter · CH</i>"]
        D2["<b>02 Founder</b><br/>Real people,<br/>no PEP / sanctions<br/><i>Specter · OS</i>"]
        D3["<b>03 Lead investor</b><br/>Lead is <i>actually</i> leading<br/>in last 60 days<br/><i>Specter signals</i>"]
        D4["<b>04 Round dynamics</b><br/>Size, valuation,<br/>pro-rata math<br/><i>Specter · GPT-5.5</i>"]
        D5["<b>05 Mandate</b><br/>LPA · IC · signing<br/>matrix · pure code<br/><i>MANDATE.md</i>"]
        D6["<b>06 Wire safety</b><br/>BEC · shell · sanctions<br/><i>WHOIS · DKIM · OS · CH</i>"]
    end

    V{{Verdict layer<br/>any block ⇒ HOLD<br/>any flag ⇒ REVIEW<br/>else PROCEED}}

    P --> D1 & D2 & D3 & D4 & D5
    SPA --> D4
    WI --> D6
    D1 & D2 & D3 & D4 & D5 & D6 --> V
    V --> M["IC Memo<br/>(deterministic + LLM lede)"]
    V -- proceed --> L["Mock wire ledger"]
    V -- hold --> A["Amendment PR<br/>(Cursor Composer 2)"]

Each desk emits its own desk.start → desk.citation* → desk.resolved event sequence. The orchestrator never blocks on the slowest desk — Promise.allSettled plus a 30-second per-desk hard timeout guarantees the verdict layer always runs.

The amendment loop — the agent edits the agent

flowchart LR
    BLOCK["Wire-safety BLOCK<br/>e.g. lookalike acrne.co"]
    OVR["GP clicks<br/>'Override & amend'"]
    OCTX[("OverrideContext<br/>+ runId + clause")]
    COMP["Cursor SDK<br/>Composer 2"]
    MD[("MANDATE.md")]
    DIFF["Unified diff<br/>+ rationale paragraph"]
    PR["Draft PR<br/>(Octokit · optional)"]
    REVIEW["Managing partner<br/>reviews like code"]
    MERGE[("Mandate v + 1<br/>future runs enforce<br/>the new clause")]

    BLOCK --> OVR --> OCTX --> COMP
    MD --> COMP
    COMP --> DIFF --> PR --> REVIEW --> MERGE
    MERGE -. enforced by Desk 05 .-> BLOCK

Every override is a learning event. The mandate becomes the fund's compounding moat — versioned, reviewed, merged.

Quick start

The backend runs end-to-end against fixtures with zero API keys required.

# Backend (port 3001)
cd backend
npm install
npm run dev

In a second terminal — the smoke driver POSTs both seeded scenarios to /api/run, prints the streaming desk events, fetches the memo, and draws the amendment-PR draft for the BEC run:

cd backend
npm run smoke

Frontend (port 3000):

cd front-end
npm install
npm run dev

Then open http://localhost:3000.

Backend API

All endpoints are versionless and consumed by the UI over HTTP.

Method	Path	Description
`GET`	`/api/health`	Liveness probe
`POST`	`/api/run`	Start a diligence run; streams `RunEvent` lines as SSE (`text/event-stream`)
`GET`	`/api/memo/{runId}`	Returns `MemoData` for a completed run
`POST`	`/api/amend`	Drafts an amendment PR from an `OverrideContext`

The full SSE event schema lives in backend/lib/contract.ts — a single source of truth shared by backend and UI.

Curl smoke (click to expand)

# Health
curl -s http://localhost:3001/api/health

# Streaming run — clean Acme (expected verdict: PROCEED, 6/6 desks pass)
curl -N -X POST http://localhost:3001/api/run \
  -H 'Content-Type: application/json' \
  -d '{
    "prompt": "Wire $2,000,000 to Acme Robotics for their Series A. Lead is Sequoia. 50% pro-rata of our $4,000,000 allocation.",
    "files": [
      {"name":"acme_spa.pdf","mime":"application/pdf","size":0,"ref":"spa"},
      {"name":"wire_instructions_clean.pdf","mime":"application/pdf","size":0,"ref":"wi-clean"}
    ],
    "fixtureSeed": "clean-acme"
  }'

# Streaming run — SPECTER_FLOW canonical Dex deal (expected verdict: REVIEW)
curl -N -X POST http://localhost:3001/api/run \
  -H 'Content-Type: application/json' \
  -d '{
    "prompt": "Wire $2,650,000 to Dex for their Seed. Lead is Andreessen Horowitz. 50% pro-rata of our $5,300,000 allocation.",
    "files": [
      {"name":"dex_spa.pdf","mime":"application/pdf","size":0,"ref":"spa"},
      {"name":"wire_instructions_clean.pdf","mime":"application/pdf","size":0,"ref":"wi-clean"}
    ],
    "fixtureSeed": "dex-meetdex"
  }'

# Memo for a completed run
curl -s http://localhost:3001/api/memo/<runId>

# Amend (draft a PR after a wire-safety BLOCK)
curl -s -X POST http://localhost:3001/api/amend \
  -H 'Content-Type: application/json' \
  -d '{
    "runId": "<runId>",
    "blockingDesk": "wire",
    "blockingReason": "Lookalike domain acrne.co vs verified acme.co — wire_safety §6.2",
    "clause": "wire_safety §6.2",
    "rationale": "Confirmed BEC pattern; tighten policy."
  }'

Demo scenarios

The demo ships with two seeded buttons — one contrast, one story.

Scenario	Verdict	What the audience sees
🟢 Clean Acme deal	`PROCEED` · 6/6 desks pass	Tiles light up green over ~30 s · IC memo renders · wire queues
🔴 BEC Acme deal	`HOLD` · Wire desk BLOCK	Five tiles green; Wire safety lands red with: lookalike domain (`acrne.co` ↔ `acme.co`, edit distance 1), domain age 6 days, DKIM fail, beneficial-owner mismatch · BLOCK modal cites `wire_safety §6.2` · GP clicks Override & amend → Cursor Composer 2 drafts a `MANDATE.md` PR
🟡 Dex (`meetdex.ai`)	`REVIEW` · Founder desk flag	Five tiles green; Founder desk turns yellow with two SPECTER_FLOW.md §4 flags surfaced: `founder_departed_before_close` (Harry Uglow exited 26 days before the Seed close) and `ex_founder_now_at_investor` (his current tagline references "a16z speedrun scout", a listed investor). Verdict is REVIEW — the data does not block the deal, but the AI underwriter has surfaced what a human reader might miss.

Set DEMO_FORCE_FIXTURES=true to skip every live API and serve from backend/fixtures/. The on-stage panic button.

Project layout

underwriter-cursorhack/
├── backend/                          # Next.js 16 Route Handlers (port 3001)
│   ├── app/api/
│   │   ├── run/route.ts              # POST · SSE stream of RunEvent
│   │   ├── memo/[runId]/route.ts     # GET · MemoData
│   │   ├── amend/route.ts            # POST · AmendmentDraft (Composer 2)
│   │   └── health/route.ts
│   ├── agents/
│   │   ├── orchestrator.ts           # fans out 6 desks, joins at verdict
│   │   ├── parse-prompt.ts           # GPT-5.5 → ParsedDeal (Zod)
│   │   ├── synthesise.ts             # findings → verdict (pure)
│   │   ├── memo.ts                   # findings + verdict → MemoData
│   │   ├── amend.ts                  # override → AmendmentDraft
│   │   └── desks/                    # company · founder · investor
│   │                                 # round · mandate · wire-safety
│   ├── lib/
│   │   ├── contract.ts               # ⭐ THE SEAM — shared types
│   │   ├── mandate-loader.ts         # gray-matter on MANDATE.md
│   │   ├── mandate-evaluator.ts      # pure rule evaluation (no LLM)
│   │   ├── ledger.ts                 # mock wire ledger
│   │   ├── cache.ts                  # in-memory + fixture fallback
│   │   └── sources/                  # specter (SPECTER_FLOW.md impl)
│   │                                 # companies-house · opensanctions
│   │                                 # whois · pdf-parse · llm
│   ├── fixtures/
│   │   └── specter/snapshots/        # canonical SpecterSnapshot fixtures
│   │                                 # (dex.json, acme.json)
│   ├── MANDATE.md                    # the policy file the agent runs against
│   └── docs/                         # Backend.md · ARCHITECTURE.md · DEMO.md
└── front-end/                        # Next.js 16 UI (port 3000)
    └── app/
        ├── components/               # MandateScreen · RunScreen · MemoScreen
        │                             # DeskTile · VerdictBar · BlockModal
        │                             # AmendmentPR · CreatePRModal
        └── state/                    # types · initial · fixtures

Configuration

All external APIs are optional — the backend defaults to fixtures and runs cleanly with zero keys. See backend/.env.example for the complete list (Cursor SDK, OpenAI, Specter, Companies House, OpenSanctions, WHOIS, GitHub).

DEMO_FORCE_FIXTURES=true   # bypass every live API — on-stage panic button
DEMO_GITHUB_REPO=          # if set, amendments open real PRs via Octokit

Design principles

The mandate is the spine. Every decision is grounded in MANDATE.md. No agent has authority outside what the mandate grants. Overrides become amendments via PR.
Six desks, parallel by default. Each desk is a single-purpose subagent with one job, one data-source family, one output shape. They don't talk to each other — they join at the verdict step.
Load-bearing data, not decorative. Every desk has a primary source it cannot function without. If the source is down, the desk reports degraded confidence rather than fabricating.
Calibrated escalation. Desks don't say "looks fine." They say PASS, confidence 0.94, basis: [Specter ID, Companies House filing, comparable round]. The verdict layer treats confidence as input, not noise.
The agent edits the agent. Every override drafts an amendment PR. The fund's playbook compounds in git.
Honest failure. Three tiers (graceful degradation → desk-level flag → orchestrator error). Citations carry cached: true when fixtures fired. We never fabricate a finding. We never silently omit a desk.

What this deliberately does not do

No real money movement (mock ledger only) · no email sending (drafts only) · no authentication (single-tenant demo) · no run persistence beyond process memory · no automatic retries (one shot, then fixture) · no PII logging.

Credits

Built for the Cursor Hackathon. Thanks to:

Cursor — the SDK, Composer 2, and the cloud sandboxes that make subagents real
Specter — the load-bearing dataset for company, founder, investor, and round-dynamics desks
OpenAI · Companies House · OpenSanctions — the rest of the data spine

_{UnderWriter — your fund's policy, executed.}

Name		Name	Last commit message	Last commit date
Latest commit History 36 Commits
.github/workflows		.github/workflows
backend		backend
front-end		front-end
.gitignore		.gitignore
AGENTS.md		AGENTS.md
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

UnderWriter

Autonomous VC Diligence, in 60 Seconds

The pitch in one paragraph

Why this matters

Technology stack

System architecture

The diligence run, end to end

The six desks at a glance

The amendment loop — the agent edits the agent

Quick start

Backend API

Demo scenarios

Project layout

Configuration

Design principles

What this deliberately does not do

Credits

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

UnderWriter

Autonomous VC Diligence, in 60 Seconds

The pitch in one paragraph

Why this matters

Technology stack

System architecture

The diligence run, end to end

The six desks at a glance

The amendment loop — the agent edits the agent

Quick start

Backend API

Demo scenarios

Project layout

Configuration

Design principles

What this deliberately does not do

Credits

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages