Intelligence for Existing CCTV Systems

Most organizations have CCTV and have invested in surveillance. Yet teams still spend hours—or even days—reviewing footage before making a decision. The cameras work; humans must connect events across time, locations, and multiple screens. Our AI Investigator unlocks the intelligence hidden inside that investment. Operators can ask: What happened? Is this normal? Who was involved? Does it require action? The system investigates the footage and returns a verified answer with video evidence. The result is a new way to operate CCTV. Investigations that once required hours or days of manual review can now be completed in minutes—and sometimes seconds. This is not merely faster video search. It is a paradigm shift from passive recording to an intelligent system that understands the physical world and delivers verified answers. The platform builds a camera graph enriched by VLM scene understanding, allowing it to learn how spaces connect and what each location represents. RT-DETR detects people and objects, BoT-SORT tracks movement, and OSNet links the same person across 28 cameras under one identity. Gemma converts scenes into searchable activity logs and recognizes fights, package handling, unusual crowding, and routine logistics. When an investigation is needed, the system reconstructs a person’s journey across cameras, summarizes the movement, and provides relevant clips. A second vision model verifies each result before delivery, reducing false positives and unnecessary alarms. The complete pipeline runs on a single AMD Instinct MI300X using ROCm, MIGraphX, and vLLM-ROCm. Its 192 GB of HBM3 allows detection, tracking, re-identification, vision understanding, and agentic reasoning to operate together across 28 streams. No new cameras. No infrastructure replacement. Just faster investigations, better decisions, and more value from existing systems. We are not building AI that watches cameras. We are building the intelligence layer for the physical world.