Video Intelligence Agent
AI agent — hours of video → scene detection for investigations & court. Used by Govt. of Canada & Singapore; outperformed 12 labs.
Research · Video intelligence · Government use
Hours of video → the scenes that matter
A Video Intelligence Agent that analyzes long-form footage to identify particular scenes — built for criminal investigation and court workflows. Used with the Government of Canada and Singapore, and evaluated as outperforming twelve competing labs.
Hours→scenesLong-form video
GovCanada · Singapore
12 labsOutperformed
CourtInvestigation-ready
System End-to-end video intelligence pipeline: ingest long recordings → hierarchical encoding (frame → shot → scene → case context) → retrieve / rank scenes matching investigative queries → human-verifiable outputs for case files.
Problem Investigators and legal teams drown in CCTV, bodycam, and case video. Finding a specific event across hours of footage is slow and error-prone when done only with linear human review. Prior lab systems often lacked hierarchy, auditability, or throughput for real case loads.
Our contribution Designed a deep networked agent with explicit hierarchy, layering, data mapping, and pipelining — so scene search is structured and scalable rather than a flat embedding sweep. Mapped raw video signals into searchable scene/event records suitable for investigation and court use.
Method
| Layer | Role |
|---|---|
| Hierarchy | Multi-level representations — frame → shot → scene → case |
| Layering | Stacked perception and reasoning (detection, temporal linking, semantics) |
| Data mapping | Raw signals → structured scene / event records |
| Pipelining | Ingest → encode → retrieve → verify at hours-scale throughput |
Achievements
- Deployed / used in government contexts (Canada, Singapore)
- Benchmarked ahead of twelve labs in comparative evaluation
- Practical path from bulk video to investigation- and court-usable scene intelligence