Snakes kept getting onto a large site, right where people walk at night, and one missed snake can mean a bite. The site already had 70 cameras, but no one could watch every screen all night.
We built a computer vision system that watches every feed, confirms each sighting and alerts the site team as soon as it's spotted.
On a 1920 by 1080 camera, a snake is about 70 by 40 pixels, often at night and half in shadow. A standard object detector downscales the frame to 640 pixels before inference, and the snake shrinks to a 23 by 13 smear. False positives are the opposite problem. Leaves, shadows and cables move all night, and at two frames a second across 70 cameras, the system screens about 12 million frames a day. If every leaf sends an alert, people stop trusting the alerts.
Each feed is sampled twice a second. Motion detection runs first: background subtraction compares every frame with the camera's learned background and passes on only the regions that changed. Object detection comes next. The detector runs on each changed region cropped from the full-resolution frame, so the snake keeps every pixel, and a detection only counts once it holds for five consecutive frames. Validation is last. A vision-language model reviews the saved frame with the whole scene around the box and returns snake, not a snake or not sure. Only a snake verdict triggers an alert, sent to Telegram with the frame, camera, time, confidence and the model's reasoning.
Running all 70 cameras on a single GPU server was the last hard part. Seventy feeds at two frames a second is 140 frames a second to screen. Motion gating keeps the detector on changed regions only, at 6 to 8 milliseconds per inference on the GPU, against 13 to 15 on a CPU. Cameras are assigned round-robin across inference workers and motion analysis goes to whichever thread is free, so no worker is overloaded and no checks are dropped. Exclusion masks cover the timestamp burned into each feed, which background subtraction otherwise reads as constant motion. We tested it against the site's full camera setup.
The same technology can be adapted to other use cases, like catching intruders in restricted areas at night and tracking objects across a factory floor. It runs on the cameras that factories, warehouses and construction sites already have.
Explore how our core modules are deployed across different operational architectures.
1,500 dormant songs became a content factory: 800 lyric videos in two months, 100+ music videos in six weeks.
A 3D Krishna with voice, memory, and guardrails that hold character under attack. 50,000 interactions, ~3 seconds each.
Editorial agents catch beats, self-hosted presenters deliver them, one operator supervises. Broadcasting for people who will never watch the 8pm news.
