The Problem: Data Poverty in Venture Capital
Startup equity is one of the most consequential financial structures in technology, yet it remains one of the least accessible. A typical seed-stage company has a cap table managed in a spreadsheet that only 1–2 people understand. By Series A, the math becomes genuinely complex — liquidation preference waterfalls, anti-dilution ratchets, option pool shuffles, and conversion mechanics create a web of interdependencies that even experienced CFOs find error-prone.
The Three-Persona Gap
The friction compounds because cap table stakeholders have fundamentally different needs:
- Founders: need full-table simulations — fundraising dilution, exit distributions, anti-dilution scenarios.
- Investors: need personal return analysis — liquidation preference outcomes, SAFE conversion math, pro-rata allocations.
- Employees: need simple answers — "How many options have vested? What are they worth?".
The Solution: Intent, Not Math
Dropping cap table data into ChatGPT produces plausible-sounding nonsense. General-purpose LLMs hallucinate share counts, fabricate liquidation preference stacks, and silently ignore anti-dilution mechanics.
CapTable Copilot is an AI-native equity modeling assistant built around a single architectural insight: the LLM should understand intent, not perform math.
Three Execution Layers:
- Conversational AI (Gemini 2.5 Flash): Handles natural language understanding, persona detection, parameter gathering, and result interpretation. It never performs financial calculations directly.
- Precision Math Engine (FastAPI): A Python backend handles all financial simulations using "precision" arithmetic. Every calculation is deterministic and auditable.
- Interactive Presentation (React): Renders structured data from tool calls, letting users explore "what-if" scenarios without additional LLM round-trips.
Persona Workflows
The system dynamically adapts its behavior based on who's talking, creating strong security boundaries and personalized logic.
For Founders
Full cap table analysis suite. Everything feeds from the master cap table document.
- Fundraising Dilution
- Exit Waterfall Simulation
- Down Round & Anti-Dilution
- Option Pool Refresh
For Investors
Standalone personal tools that map return paths — no master cap table required.
- Investor Return Calculator
- SAFE/Note Conversion
- Pro-Rata Rights Calculator
For Employees
Grant-level analysis tools enforcing security boundaries isolated from sensitive cap table data.
- Equity Status Dashboard
- Exit Payout Estimator
App Showcase
Landing Screen
Full Cap Table Analysis
The Insight
LLMs are orchestration engines, not computation engines. The moment we stopped trying to make the model do financial math and started making it route to the right calculator with the right inputs, the product became dramatically more reliable.
Architecture Stack
- Frontend: Next.js 15, React 19.
- AI: Vercel AI SDK v6, Gemini 2.5 Flash, Prompt Engineering,
- Math Engine: Python FastAPI, ETL, Pydantic v2, LlamaParse OCR.
- Datastore: PostgreSQL via Prisma ORM for capTableState.
Technical Execution
Tools split into Backend-computed (e.g. waterfall) requiring Decimal precision and Client-side tools executed instantly in-browser.
Implemented server-side data injection to prevent LLM hallucinations from silently breaking UUID cross-references during state reconstruction.
Cap table truncation, session facts extraction, and history trimming ensure large datasets fit well within context limits.