How it works
Intelligence at the edge.
Privacy by architecture.
Most vision vendors stream your cameras to their cloud. We do the opposite: the entire pipeline — decode, inference, decision, storage — runs on a server inside your premises. The only thing that leaves is the alert you choose to send.
Connect
We attach to your existing IP cameras over RTSP. Each camera gets its own resilient worker that reconnects automatically — a dropped stream never takes down its neighbors.
Understand
A micro-batched inference engine shares each GPU across hundreds of streams. Temporal voting confirms events across frames before anything is called an alert.
Act
Verified events flow to your dashboard, control room, or ERP over standard webhooks. Every event is journaled locally first — if your network drops, nothing is lost.
TEMPORAL_VOTING
Events are confirmed across consecutive frames and tracks before alerting. The result: alarms your team actually trusts.
MICRO_BATCHED_INFERENCE
One GPU, hundreds of cameras. Streams are batched dynamically so idle channels never waste compute.
OFFLINE_OUTBOX
Every detection is written to a local journal before delivery. Network outages delay alerts — they never delete them.
SELF_CALIBRATION
Per-camera exposure and sensing auto-tune over a 24-hour cycle, adapting to lighting, weather, and lens drift — your tolerance thresholds stay exactly where you set them.
VLM_VERIFICATION
High-stakes alerts are double-checked by an on-premise vision-language model — and the check fails closed.
CONFIG_AS_CONTROL
Every module is governed by one human-readable config file. No hidden state, no vendor lock screens.
Cameras engineered to run on a single server
Continuous operation, lights-out reliable
From event to verified alert on-premise
Of footage stays inside your walls