START WITH THE EVIDENCE
Explore three episodes.
Recovery, a test alert without recovery, and seated closure. Explore body pose alongside saved observations, decisions, applied commands and transmission changes. Recorded and replay decisions are identified separately.
Public access · No key needed. The review contains no camera images, video or audio. Playback does not run new inference.
CARE-Vision selects what to observe next. Vision and the deterministic controller run on AWS Graviton with OpenCV 5 + COOL; the Edge captures, samples and returns the requested evidence. Explore the illustrated workflow below.
Measured benchmark · 18 replays · 8/9 paired comparisons passed
Fixed and adaptive transmission
25 October 2026 · Fixed 5 FPS versus adaptive 2/5 FPS. Both policies run person vision and incident rules on AWS Graviton with OpenCV 5 + COOL.
| Recording | Less JPEG data | Less cloud CPU | Pairs passing all gates |
|---|---|---|---|
| Recovery | 53.54% | 53.12% | 3/3 |
| No recovery | 53.47% | 53.07% | 2/3 |
| Sitting | 47.17% | 46.61% | 3/3 |
Retained failure: one adaptive no-recovery run reconnected and missed its expected test alert. Of 171 locally completed sends, 160 were received and processed. Complete outcome equivalence is not established.
Sitting stayed UNRESOLVED without alerts in both arms, a known limitation. The person-only profile disables pose, ROI and seated closure in both arms. The fixed arm uses a declared logical-mode adapter without physical sensing changes or fabricated ACKs.
CPU uses direct process counters; JPEG excludes transport overhead. The experimental additional-delay margin is two seconds. Known recordings and repeated runs do not establish general fall accuracy, a 30 FPS comparison or billed AWS savings.
Standard OpenCV 5 versus COOL
This is a separate experiment: identical-blob forward medians were 67.701 ms standard and 67.733 ms COOL, with 1,920 exact tensor/detection comparisons. No COOL speedup was demonstrated. Original full-pipeline output equivalence failed in 102/240 frames.
Saved results, no inference started by this page. Full methods, source/input hashes, ranges and failures are supplied in the Devpost award evidence attachment.
SYSTEM EXPLORER
See the loop in action.
Follow a frame from capture to the next agent decision.