Solutions
YOLO + agent — small video PoC
For Solutions I want a small idea, not a platform pitch: YOLO plus an agent as a simple video PoC.
I tried it the humble way on my OpenClaw setup — asking YOLO to find people in photos: famous faces, or even myself. About fifty-fifty unless the picture is obvious. Unclear shots are hard. Useful honesty. Not magic.
You don’t have to use it for everything. But you can use it more for PoCs and test projects.
I’ve also tried simple counting and spotting jobs: counting people, counting goods, recognising a person in frame. Same lesson — it works when the scene is clear enough; it struggles when it isn’t. That’s exactly why a small PoC is valuable before anyone buys a stack of cameras and a full analytics programme.
The old VA pattern was heavy: many samples, long test runs, sometimes buying lenses first.
With YOLO you can flip it: collect some footage first, ask what is on the feed, then put an agent on that description — flag what looks abnormal, keep a timeline of frame captures, store those readings. When something happens, jump back to the clip before and after that timestamp.
Pros: cheap to start; forces a narrow question; early truth about cameras and false positives; tests whether “next step” automation is useful; builds a retrieval trail.
Cons: demo scenes lie; lighting and angles break you; alerts get noisy; agents can mis-route junk; a cute PoC is not production; face or object ID can be hit-or-miss.
Small loop: detect, interpret, act. Enough to decide whether to scale — not a shortcut past design.