Edge AI VideoIntelligence Platform

It's an AI-powered video investigation platform that turns raw camera footage into searchable incidents and AI-generated operator reports. An edge box next to your cameras runs real-time object detection, tracking, and zone rules, then a compact vision-language model triages every event on-device — asking, in effect, 'would a guard look twice?' Only events worth attention are escalated to a frontier cloud model for deep analysis, which produces a summary, a risk level, and a recommended action. Everything becomes searchable in natural language.

No — it sits above them as the reasoning layer. Your cameras, video management system (Avigilon, Milestone, Genetec, or others), and recording infrastructure stay exactly as they are. The platform ingests standard RTSP streams, which every IP camera exposes, so connecting real cameras requires no new hardware and no rip-and-replace. Most security teams drown in alerts; we take the alert, the footage, and the context, and return a reasoned recommendation the way an experienced operator would.

It's a filter-and-refine architecture in three stages. First, a YOLO-family detector with ByteTrack tracking finds people and vehicles and applies your zone rules in real time on the edge box. Second, a compact vision-language model (such as Gemma or Qwen-VL) reviews keyframes on-device and dismisses routine activity — those events are logged and audited but never sent to the cloud. Third, only escalated events go to a frontier model like Gemini or Claude for deep analysis. The triage fails open: if the edge model is ever unavailable, events escalate rather than being dropped, so an outage never loses an event.

Privacy is architectural, not a policy checkbox. Full-resolution video never leaves your site — it stays on the edge box and your existing recording infrastructure. When an event escalates to cloud analysis, only a small set of downsampled keyframes is sent, and routine events are handled entirely at the edge with no cloud contact at all. Every incident carries a full audit trail recording which model analyzed it, where that model ran (on-device or hosted), what images it saw, and what it concluded.

Every analyzed incident is embedded into a vector index (PostgreSQL with pgvector), so operators can search past footage the way they'd describe it to a colleague — 'person loitering near the loading dock after hours' — instead of scrubbing timelines. Searches combine semantic similarity with structured filters like risk level, event type, camera, and time range, returning ranked incidents with their clips, keyframes, and AI reports attached as evidence.

A proof of concept on your recorded footage — detection, triage, AI reports, and the investigation console — typically takes 4-6 weeks. A live pilot with RTSP camera ingest and an edge box on-site usually lands in 8-12 weeks, and fleet rollout across sites follows from there. Because uploaded video and live cameras share one code path, everything validated in the POC carries directly into production.