04 · Computer vision
Road Defect AI
Hands-on dataset and model work for road defect detection, including a custom tool built because the review workflow needed to be faster and more consistent.
Real model work
Model quality begins with the data humans prepare.
My experience is in the difficult, repetitive side of applied computer vision: preparing and reviewing training material, correcting detections, improving annotation consistency, checking false positives and false negatives, and iterating the dataset instead of treating training as a single button press.
The imagery is real road data, with all the ambiguity, lighting variation, surface differences and edge cases that come with field conditions.
Purpose-built tooling
A faster correction loop.
I built a Flutter YOLO annotation editor to improve the workflow. It supports dual-pane image review, YOLO text files, severity classification and rapid correction of model output.
Building the tool connected software design directly to model quality: reducing friction in the review process makes it more practical to maintain consistent annotations and spend time on the decisions that actually affect the dataset.
Engineering view
AI output is evidence to evaluate, not truth to accept.
This project shaped how I use AI more broadly. Convincing output can still be wrong. Good engineering depends on validation, representative tests, traceable changes and a willingness to correct the tool when the evidence disagrees.