Tejas Datla

Edge Confidence Router is an on-device camera reader that knows when it doesn’t know. Meta’s Perception LM 1B runs on an 8 GB Jetson Orin Nano and reads numbers and codes from images in ~1.2 s. The router scores each answer from the model’s own token probabilities (the probabilities of the digits it wrote, multiplied together) and keeps it on the device only if it clears a threshold of 0.80, chosen once on a separate dev set and frozen before any test. For IDs with a known format, the answer must also match the format (five digits for British Railways locomotives). Everything else goes to Meta’s Muse Spark in the cloud. With no connection, the device still answers immediately, marks unsure answers as unverified, and queues those frames; when the link returns, Muse Spark verifies the queue: read now, ask later. No training or fine-tuning. Results (rules and labels committed before each run, each test run once): Locomotive numbers (25 Wikimedia Commons photos): 9 kept on the device, all 9 correct; the format check caught the one confident mistake that confidence alone would have kept; 21/25 after sync. Jersey-number benchmark (25 held-out photos): router 24/25, matching sending everything to the cloud within noise, with 11 answered on the device. Offline: 22/25 immediately, 23/25 after sync. Long label codes are beyond the 1B model; the router sent all 9 to the cloud and kept none wrong. A live dashboard shows each frame’s answer, confidence, route (on-device, Muse Spark, or queued) and the sync log. Every result file, threshold and rubric is in the repo, and the threshold rule can be re-checked without a Jetson.