Adam Burich

A drone watches highway traffic. The system tracks every vehicle, works out the road geometry from nothing but the traffic itself, and scores each driver 0–100 on how dangerous their behavior actually is — tailgating, forcing others to brake, cutting into gaps, weaving — instead of crude proxies like raw speed.

The part I was least sure about was whether the number meant anything. A score is easy to produce and easy to believe, so I put it in front of people: raters watched blind pairs of clips and picked the worse driver, and I measured the score against their choices instead of my own judgment.

Every optimization shipped only after a bit-identical check against the version it replaced. A scoring rework that failed to improve human agreement was shelved with its analysis instead of merged. The open gaps — calibrating the thresholds rather than just the ranking, per-rater weight profiles, curved-road geometry — are written down, not glossed over.

Danger Index — the system's output

92%

Human–system agreement on held-out research recordings

88.5%

Agreement on real drone footage, full pipeline, never seen during development

81%

Of held-out conflicts where the score was already rising beforehand

2–4 ms

Per frame scoring on a commodity CPU — real time with headroom

3 weeks

Start to finish, solo. ~25,000 lines, 257 commits

Python · PyTorch · YOLOv8 · OpenCV · TensorRT · OR-tools · NumPy
Provisional patent drafted — 20 claims, 10 figures