§0problem
← All case studies

A design validator for a banking app

Level 4 · 2026 · for a banking software vendor

Many small UI bugs had three possible origins: the implementation, an inconsistent design file, or the analysis. Nobody could tell which one without hours of comparison.

§1built

What was built

A design validator that finds the origin. It is a deterministicGives the same result each time, with no language model. pipeline. Same input, same findings file, byte for byte No language model runs inside the rule engine. Model work lives in skillsA file that tells an agent how to do one task. In my method, a skill applies a rule. It does not make one., where an agentAn AI program that writes and changes code, for example Claude Code or Cursor. On this site, "agent" means a coding agent. writes claims and code checks them.

§2taught

What it taught

  • The validator reports and never changes code. Another agent changes the code. That separation is the safety property.
  • A missing finding is better than a false one.
  • A single quality score was rejected. A score can read perfect while blocking findings sit under a baseline.
  • A blocking gate was rejected. False blockers burn the agent's limited fix attempts. The validator reports and routes.
  • The design file is evidence, not authority. When file and library disagree, the tool reports both values and merges nothing.
§3open

What is still open

About 13% signal at first contact with real files; precision and recall of 1.0 on the tuning corpus only
Review chose the agent-with-validator arm in every pair that produced a pull request (6 ticket pairs)
6 design-tool API calls a month on a view-only seat; retry-after about 4.4 days