Your model on your own hardware

Models trained on your verified footage are yours: download the weights and the dataset from the model view. There is no per-image inference fee attached to a model you run yourself.

Distilled models are sized for edge hardware, next to the cameras. Footage is processed where it is captured; what leaves the site is the answer, a detection, a count, an alert, not the video.

tip · Plan for the enclosure, not the benchmark. Sustained load in a sealed box throttles; a swapped camera changes the image in ways the model notices before you do; and video decode, not the network, is often the real bottleneck. Budget headroom for all three.

The same model runs wherever you draw the boundary: your VPC, your on-prem cluster, or LexData-managed infrastructure. The loop behaves identically in all three.

When the loop retrains, what moves to your deployment is an updated model, not your footage moving to us. Your streams stay on site; what the loop keeps is the review queue - the flagged evidence frames and the verdicts made on them - which is exactly the material the next training run needs.

Enterprise includes a deployment and security review, with our engineers working through your infrastructure and compliance requirements before anything ships. Start that conversation at reach@lexdatalabs.com.