I built a system that detects community infrastructure issues like potholes, road cracks, and debris using video feeds from vehicles.

The pipeline processes dashcam/autonomous vehicle footage, extracts frames, runs CV based detection, and aggregates results over time to reduce noise. Each detection is mapped with metadata to help visualize problem areas on a simple dashboard.

The hardest part wasn’t the model itself it was iterating quickly. Small changes in preprocessing, thresholds, or tracking logic required long reruns of video pipelines, which slowed down experimentation.

That’s where **RocketRide **helped. It significantly improved my iteration speed by reducing the friction between making a change and testing it. I could quickly validate preprocessing tweaks, experiment with detection thresholds, and debug intermediate outputs without constantly rebuilding the pipeline.

This led to faster experimentation cycles, better detection stability, and more time spent improving the actual logic instead of dealing with tooling overhead.

Overall, the biggest gain wasn’t just in building the system but in how quickly I could iterate on it.

If you’re working on similar pipelines or dev heavy ML workflows, definitely check out RocketRide https://rocketride.org/