DriftRadar

DriftRadar is a web + desktop app (agent) that monitors robot perception pipelines in production and flags “silent failures” like sensor misalignment, lens contamination, lighting shifts, domain drift, and degraded detections. It plugs into ROS/ROS2 topics or log files and computes lightweight health metrics (feature distribution shifts, detection confidence entropy, reprojection residuals, timestamp jitter) without needing full retraining. Teams get a dashboard with trend lines, incident timelines, and automated “what changed” hints tied to deployments, environment, and sensor telemetry. It also generates a minimal reproduction bundle (short clips + metadata + config snapshot) so engineers can debug quickly and share with vendors. This is an AI app + traditional app combination: AI helps summarize incidents, cluster failure patterns, and propose likely root causes, but the core value is reliable monitoring and alerting. Reality check: it won’t magically fix bad models; it just makes failures visible early.