Overview
When a disaster wipes out a business, someone has to prove the income it lost. DeepWell builds that number and helps defend it in court. Most of the heavy lifting is automated, but AI canโt commit anything on its own: a named expert has to sign the result and defend every figure in a deposition. So the product depends on people stepping in at the right moments to check a decision, make a judgment call or sign off.
As the only designer on the team, I redesigned v1 end to end and built it in code.
DeepWell, 2026 (in active development)
Context & Scope
DeepWell was built by a group of advisors who do this work today, one claim at a time. The engine already existed when I joined, but the user experience was lacking.
The core challenge was maintaining density and trust at the same time. Every claim sits on a deep stack of documents, rules, calculations, and AI reasoning. All of that powerful backend was visible in the front end when I joined. My job was to let the automation fade into the background while making every human checkpoint clear and deliberate.
Key Decisions and Tradeoffs
The app opens on what needs a person and lets everything settled recede. The AIโs work shows up as outcomes rather than process, and the heavier a human decision is, the more deliberate the interaction becomes.
The complexity of the product mirrors the complexity of what an analyst is working on at any given time.
The Product
DeepWell is a web app in dark mode grayscale with a bright orange saved for whatever matters most on the screen. Built in React, Vite and Tailwind with Claude Code.
What's Next
v1 of DeepWell only scratched the surface, and that's all I can say for now!





