DriftRadar

DriftRadar is a web + edge desktop app that continuously monitors computer-vision models running on robots and flags real-world performance drift before it becomes downtime. It passively samples frames and metadata from the robot pipeline, computes lightweight embeddings, and tracks distribution shifts (lighting, motion blur, camera focus, seasonal changes, new parts). When drift crosses thresholds, it generates a short “what changed” report, suggests targeted data to collect, and produces a prioritized re-labeling queue. It also validates camera health signals (exposure, sharpness, dropped frames) to separate sensor issues from model issues. This is not a full MLOps platform; it’s a focused “early warning system” that plugs into existing stacks via a simple SDK and can run offline on factory networks. Expect integration work and some false positives early—this succeeds only if it’s fast to deploy and earns trust with clear, actionable alerts.