VC GROUND TRUTH

A portfolio intelligence system built for a working VC investor. Deterministic ETL from board reports into Postgres, a vision-transcribed evidence layer, and natural-language answers through Claude.

PROJECT - CASE STUDY

01 · THE PROBLEM

Twenty minutes per question.

A VC partner tracks ~110 portfolio companies across board decks, KPI spreadsheets, email, and fund-admin systems. Answering one question meant finding the files, reading every tab, verifying by eye, and re-keying numbers into a fresh spreadsheet. Nothing accumulated. The next question started from zero.

02 · What I Built

One governed store. One writer. Many readers.

Sole builder, running an AI-native workflow. I scoped requirements from a verbal spec and a whiteboard sketch, designed the data model with an LLM as a thinking partner, and delegated implementation to an AI coding agent through bounded, gated prompts. Every verification, security, and infrastructure decision was mine. I didn’t hand-write the parser. I ran the process that made AI-built code trustworthy enough for a VC’s board data.

How I Worked

03 · Build Log

Four phases, all gated.

04 · Design Principle

Two lanes. One store.

05 · Automation

The Watcher Loop.

06 · In Production

Ask a question, get a board-ready answer.

The investor validated the system himself. He queried the database through Claude against a raw source file, confirmed the match, then began generating his own analyses.

EXAMPLE “How did Q2 net retention compare to what the board planned a year ago?” Answered in seconds, with the plan vintage and the actual side by side.