Onboarding a new Carta customer meant eight manual steps: emails, spreadsheets, re-uploads, and an implementation manager stitching a cap table together by hand from dozens of legal PDFs. It is now one drop-in flow, at 97% precision and 94% recall on extraction from unstructured documents. A firm arriving with 8,000 documents across hundreds of portfolio companies now onboards in weeks instead of months.
Ruchita Chandhok uses that build to work through four decisions any team putting an LLM in front of real data will face:
- Where the contract between platform and product lives. Typed schemas, versioned like APIs, not prompts. It is what lets the two teams stop blocking each other.
- What a reviewer needs to trust a machine-extracted value. Provenance on every field, down to the page and sentence. Verification in seconds beats confidence scores.
- When to build evals. Before the second feature, not after the first incident. You cannot reconstruct a regression from logs.
- What the model should never be allowed to do. Arithmetic, dedup, and date math belong in tested code. Reserve the model for judgment.
