This page reads directly from the production system — the GPU carrying the analysis, the exact models resident on it, and the totals it has indexed. What you see here is what is running.
SHA-256 of the file as received, malware scan, and a permanent evidence number — before anything else touches the footage.
Frames extracted at the analysis rate; long footage fans out into parallel segments.
Every person and vehicle found in each frame and stitched into continuous journeys across the clip.
Number plates read with ranked alternatives and confidences; appearance embeddings computed for cross-camera re-identification.
Incident rules evaluate every track — no-helmet, triple riding, zig-zag driving, abandoned objects, missing plates — as reviewable events, never verdicts.
Findings, attributes and vectors land in the evidence store; results stream to the console while the analysis is still running.
Measured: a busy 10-second clip completes end to end in ~75 seconds; a 90-second junction export in about three minutes. Raw video never leaves the deployment — frames are processed in GPU memory and only the analysis returns.
Roles are a discipline, not a decoration. Evidence-grade output is deterministic; lead-only output always requires a human decision; gated models are capability-restricted and audited. Every analysis pins the versions that produced it.
Validation is disclosed, not hidden — attribute recognition serves only the heads that pass their gates; the rest answer nothing rather than guessing.
SHA-256 of the file as received and of the stored copy, evidence numbers at intake, and a legal hold that outranks every delete.
Every AI output is a candidate until an officer accepts, rejects or marks it inconclusive. Contradictions are flagged, never smoothed over.
Sealed reports carry their chain-of-custody hash and generate the BSA 2023 Section 63(4) certificate ready for signature.
A pilot takes a day to set up. Your footage never leaves your premises.