- In development
Tire-sidewall recognition
From embossed sidewall text to structured, reviewable tire data
A client project for Radburg that reads embossed black-on-black tire markings from photographs. The pipeline extracts size, load and speed indices, DOT markings, country, retread status, ply rating, brand, model and season, retaining evidence for each field so an operator can review uncertain results.
- Python
- FastAPI
- PaddleOCR
- llama.cpp
- React
- TypeScript
Document pipeline
- Unwrap the sidewall into a polar image and split it into sector crops.
- Combine OCR from multiple views with a self-hosted vision-language model pass.
- Validate candidates against size grammar, load and speed tables, and DOT rules.
- Return structured JSON with field-level provenance and acceptance or review status.
Production notes
- Current scope
- A working web deployment with ongoing accuracy and delivery work. A native Android app, persistent queue and client GPU rollout remain delivery milestones; speed and precision targets are not presented as achieved results.
Outcome
A reviewable extraction workflow for difficult industrial imagery, combining learned recognition with deterministic checks.