Meridian

An AI investment system for venture capital firms. It learns how a fund thinks, discovers matching startups, maps the market, models returns, prepares diligence, and improves after every decision.

The core idea: a living institutional memory and reasoning system for venture investing. Not a search engine, CRM, or note taker. A fund-specific reasoning layer that connects every new opportunity to the fund's historical decisions and gets smarter with each one.

What it does

• Fund graph. Ingests prior memos, portfolio companies, rejected deals, thesis, and partner preferences. Builds an interactive graph where spatial distance means similarity.
• Conviction-driven sourcing. Finds companies that match the fund's actual decision patterns and explains why each one surfaced, including which past investments and rejections it resembles.
• Startup analysis. Company summary, historical analogues, fund-fit score, risks, competitor landscape, and missing diligence questions.
• Financial modeling. Auto-generated editable models with bull, base, and bear return scenarios. IRR, MOIC, probability of returning the fund.
• Outreach and scheduling. Founder contact discovery, personalized outreach, meeting coordination.
• Meeting copilot. Pre-meeting brief, structured live notes, contradiction detection, suggested follow-ups.
• Post-meeting output. Draft investment memo in the fund's own style, updated model, recommendation.
• Learning loop. Every approve, pass, and outcome updates Meridian. The graph and recommendation weights visibly change after feedback.
• Partner intelligence. Predicts which partners will support or oppose a deal and why.

See docs/PLAN.md for the full product plan and demo narrative. See docs/OUTREACH_AUTOMATION.md for the staged, production-safe approach to founder outreach, reply handling, and calendar scheduling.

Local development

# Terminal 1 — brain API (OpenAI + Tavily stay server-side)
cd brain && npm install && npm run api

# Terminal 2 — UI (proxies /api → localhost:8790)
cd app && npm install && npm run dev

Root .env (repo root) should include:

OPENAI_API_KEY=…
TAVILY_API_KEY=…

Vercel deployment (UI + brain API together)

One Vercel project hosts the Vite SPA and the brain as serverless /api/* routes. No separate Railway/Fly host required.

1. Import this GitHub repo into Vercel (Root Directory = repo root).
2. Framework preset: Vite (root vercel.json sets install/build/output).
3. Set server env vars (Project → Settings → Environment Variables):
   • OPENAI_API_KEY
   • TAVILY_API_KEY
   • optional: VC_BRAIN_OPENAI_MODEL (default gpt-4o-mini)
4. Leave VITE_API_BASE_URL unset so the browser calls same-origin /api/....
5. Deploy. Long sourcing/SSE runs need a plan that allows higher maxDuration (configured to 300s).

Locally you still run brain + app as two processes (Vite proxies /api → :8790). Optional separate API host still works via VITE_API_BASE_URL.

Not Supabase: live chat/sourcing needs OpenAI + Tavily compute and SSE, not a database. Supabase would only help later for durable persistence of sourced companies / feedback.