Role: Computer Vision Consultant
Automated visual inspection for sewer pipes. Crawler robots drive through underground pipes recording video; until recently a trained operator had to watch every minute of that footage looking for cracks, root intrusion, deposits, and joint defects. The system flags anomalies frame-by-frame so the operator only reviews the clips that actually need a decision.
Anomaly detection and segmentation
A per-frame model both classifies the kind of defect and segments it, so the report shows not just “crack at minute 12” but the exact pixels and their estimated extent along the pipe wall. Training had to handle the unusual visual conditions inside a pipe: severe distortion from the crawler's wide-angle lens, uneven on-board lighting, motion blur, and a long tail of rare defect types.
Scaling to 100M feet of footage
Most of the engineering effort went into the ingest and inference pipeline that processes the historical archive — over 100 million linear feet of pipe footage — without melting GPU budgets. Frames are deduplicated, sampled adaptively when the camera is moving slowly, and batched so the operator gets reviewable output within hours of upload rather than weeks of manual viewing.
Technical Highlights:
- Developed an anomaly detection and segmentation system for sewer pipes;
- Scaled it to process 100 million feet of pipe data.