The evidence platform the concept note asked for is live.

34 requirements, accounted for

Proof-of-concept scenarios: six of seven pass today. The seventh — the changed-plate anomaly — is built and test-verified; it waits only for footage in which one vehicle carries two plates. A requirement-by-requirement compliance matrix is available on request.

Why a pilot needs more GPU than a demo

Today the entire vision stack runs on one graphics processor, started on demand — the right economics for demonstrations, the wrong shape for operations, for four concrete reasons.

1 · Surge concurrency after an incident

A serious case lands as dozens of DVR exports at once. One GPU processes them serially; "searchable within one to two times its duration" needs parallel analysis workers.

2 · The 30–60 minute triage window

A GPU started cold spends its first minutes loading models — measured at 10–20 minutes. Operations need a resident GPU during working hours so the first clip of an emergency is processed in seconds.

3 · Full data sovereignty for language features

Question parsing currently uses a vetted external text service — question text only, never imagery. A tender-grade deployment self-hosts the language model, which needs one high-memory (96 GB-class) GPU and makes the platform fully self-contained.

4 · Making the models better on local footage

Helmet, face-covering and garment recognition improve by retraining on the deployment's own camera conditions — dedicated training time, separate from serving.

The measured baseline: a busy 10-second clip analyses end-to-end in ~75 seconds on one warm GPU; a 90-second junction export in roughly three minutes. Raw video never goes to the GPU provider — frames are processed in memory and only the analysis returns.