The Founder Problem
VC analysts spend their prime years copy-pasting numbers from founder decks into Excel. It's the ultimate business bottleneck. We wanted to see how fast we could break it using the "Hamstack" method.
The Moat: GPT-4 Vision
The hardest part of this build wasn't the logic—it was the input. Founders send updates as screenshots inside PDF decks. Traditional OCR fails here.
We built a multi-stage pipeline:
- Vision Processing: Every page of the PDF is converted to an image and run through GPT-4o. It "looks" at the charts and extracts the raw numerical values for ARR and Burn.
- Schema Enforcement: We wrote Zod schemas to ensure the AI only gives us perfectly formatted JSON. If the numbers don't add up, the AI flags itself for a second pass.
- Real-time Reactivity: Updates are pushed to the VC dashboard via Convex subscriptions. No
refreshbutton needed.
Why This Matters to Builders
We shipped the alpha version in 48 hours. By day 14, it was in production. In the AI era, the distance between "Idea" and "System" is shrinking to zero for those who know how to use the right tools.
