RoadAI

Problem

Transfer companies move passengers between airports and hotels. The public case study describes the problems it addresses: paper vouchers that drivers retype by hand, no record of delivery, and no visibility into vehicle use.

What I did

I defined the product, architecture and specifications, reviewed the code and the security of the design, and tested on a real device. The stack is FastAPI, Flutter, React and PostgreSQL. A Claude Haiku vision model extracts structured data from a voucher photo, and a human review step lets the driver correct it before anything is saved.

Method

• Test suites: 469 backend checks, 145 Flutter tests and 32 dashboard tests, 646 automated checks in total.
• On the first overnight real-device test the app stored very few location pings. I traced it through the phone’s own logs to a HyperOS GNSS power-saving behavior, and changed the tracking design to work around it. The real-phone re-test of that fix is still pending.

Results

646 automated checks across three suites, and a documented investigation of the HyperOS GNSS issue. The code is private; the case study contains no code or client data.

Limitations

• No production users yet; a pilot with a transfer company is in preparation.
• Extraction accuracy is not measured.
• The post-fix real-phone re-test is still pending.

AI assistance

Most of the code was written by Claude Code under my specifications and review. The decisions, test plans and the real-device investigation are mine.