- Problem
-
AI can write a financial report in seconds, but it will sometimes invent a number. In finance, one
made-up figure ruins trust in the whole report. That's the main thing holding companies back from
using AI for real work.
- What I built
-
A reporting product where every number the AI writes is checked against the source data before the
report goes out. Each figure passes four checks: is it sourced, is anything missing, is the math
right, are the names right. Numbers that can't be verified get flagged or blocked, and each report
gets a 0–100 trust score instead of a black-box pass/fail.
- Keeping it honest
-
I built a test harness that deliberately tries to trick the AI into making things up. If the trust
score drops, the build fails.
- The product call
-
I first built it to handle dozens of report types. Once the verification engine was proven, I cut the
scope to three finance reports done well: flux analysis (why numbers changed), budget vs. actual, and
price-volume-mix (what drove revenue up or down). Narrowing it was the hardest and best decision on
the project.
- How it was built
-
Solo, with AI as the developer and me as product manager. That meant a ticketed backlog, a Strategist
→ Critic → Planner review step for big design decisions, roughly 3,000 automated tests, and security
reviews.
- Built with
-
- React
- Node
- Supabase
- Claude API
- Vercel + Railway
