We don't sell a model. Brightsight is the pipeline that takes a state-of-the-art vision system, tunes it to the scene in front of it, and compiles it down to run on-prem — sub-100ms, every frame.
Person, vehicle, intrusion, abandoned object, behavior, scene change. The classifier output of a model tuned to your cameras, your lighting, your edge cases — not a generic detector.
Tracking is one of the areas where we go deepest. Multi-object tracking, re-identification across camera handoffs, persistent IDs across a site — we don't just see, we follow, frame after frame, camera after camera.
Brightsight outputs become events in the systems your operation already runs. Alarms, dashboards, control software, building systems — the platform doesn't stop at detection. It packages model output as signals your downstream stack understands, so detections turn into operational consequences without an operator in the loop.
We start with our own best-in-class detection and tracking backbones, trained on surveillance-domain footage.
We adapt the model to the customer's scene — their cameras, lighting, viewing angles, edge cases — without retraining from scratch.
We compile the tuned model down to the target hardware — x86, ARM, Jetson, whatever the rack holds — squeezing latency and footprint until it fits.
On-prem by default — where Brightsight runs best and zero data leaves the site. The customer keeps their cameras, their network, their authority. Cloud hosting is there when they need it.
Sub-100ms per frame. 24/7. Monitored, updated, and re-tuned as the scene changes.
Most vendors stop at the model. We stop at the rack.
Most vendors hand over a license and walk away. We don't. Every Brightsight deployment is anchored on hardware we install and operate. We own its health — networking, security, OS patches, model updates, inference monitoring — for as long as the system is in production.
A compiled, on-prem appliance, sized to the workload and the camera count. It lives in your rack, on your network, behind your firewall.
Inference health, network connectivity, security posture, OS updates, model versions. We see when something drifts before you do.
Security patches and model updates ship continuously. The system you deploy today is the system that gets sharper every month.
Daskell is an AI research and engineering team based in Italy. We build computer vision systems for physical security and operations — and the pipeline that makes them deployable anywhere. We work with teams that keep their data, their operators, and their decisions in the room.
Meet the team →