Everything below works right now on the live system, on a laptop or a phone. Real Hyderabad junction footage is already loaded and analyzed.
A full administrator account on the demonstration environment. It is a shared demo login — please do not load real case material; a dedicated tenant is created for any pilot.
Real footage is already there — Ameerpet X Roads, Pillar 143 Ring Road and more. Each card shows its poster and evidence number.
Open Uploads, drag any CCTV export in, set the recording date and time, and watch it move through the queue. A 90-second clip is fully indexed in a few minutes; results stream in while it runs.
Open Find and type what an investigator would ask. Counts are deterministic — same question, same answer, evidence attached.
Switch to Filters: 77 parameters — garment colours, helmet, bags, vehicle type and colour, plate fragments — plus movement direction and a drawn search area on the camera's frame.
Any result opens with its attributes, plate reading with ranked alternatives, and the reviewer decision — Accept / Reject / Inconclusive. A rejected finding stays out of sealed reports.
In the finding, choose Enhance. The classical tier sharpens real pixels and stays evidence-grade; the AI tier reconstructs detail and is banner-marked as an investigative lead only. Every operation records its hashes, parameters and operator.
The Viewer plays footage with tracked boxes on every person and vehicle, frame stepping, and a detection-density strip to jump to the busy moments. Authenticity reports the file's hash, container metadata and tamper indicators.
Pause, select a person or vehicle, choose Follow. The dossier shows every candidate sighting across cameras with component scores — appearance, plate, travel feasibility. A physically impossible hop is flagged Contradicted, never sold as a match, and each sighting takes an accept/reject decision that survives re-runs.
The Timeline tab runs up to four cameras against one shared clock — coverage bars, incident markers, and honest gaps where nothing was recording.
In Cases, create a case with the FIR number, offence, incident window and legal authority; pin findings to it as you review.
Seal a report from the pinned findings: SHA-256 chain of custody, one-click Verify, and the BSA 2023 Section 63(4) certificate rendered ready for signature. Rejected findings are listed as excluded — the audit keeps everything.
Health lists every service and the model registry — each model's version, role and disclosed limits. Command shows the supervisor's review queue; Audit shows every action, forever.
Raw video stays on the deployment's own server; the GPU processes frames in memory and keeps nothing. For a pilot, everything runs in your infrastructure or a sovereign cloud.
No, by design. Every AI output is a candidate that a trained officer accepts, rejects or marks inconclusive; contradictions are flagged; the wording is “candidate”, never identity.
The model registry discloses it: attribute recognition serves only the heads that pass validation, and the remainder are withheld rather than guessed. We would rather show the limit than have you find it.