AI plate reading across 300+ fleets and 50k+ devices: 96% recall on legible plates

Problem

Plate readability was one of the most persistent asks from customers reviewing incident video — in a claims investigation, identifying a third-party vehicle can decide the case, and that evidence goes to insurers and law enforcement. Yet reviewers were stuck scrubbing frame-by-frame to catch a plate. Just OCR the video fails twice: the real limit is source-video quality (a confident wrong read is worse than none in a claim), and a hard legal line means this can't be an automated plate reader, identify people, or become a searchable database.

Research

The reframe: the bottleneck was reading the plate, not finding it — so the decisive move was a plate-tuned OCR engine behind a swappable layer. The governing principle, drawn from legal and product both: surface only what a human could already read, and skip rather than guess — precision, not coverage, is what a claim needs.

Solution

• Fixed read quality first — moved recognition to a plate-tuned OCR engine behind a swappable implementation, so OCR models could be changed without re-architecting.
• Split detection and recognition into independently scalable pub/sub services, to bring down No. of calls to OCR models
• Shipped it as an overlay, not a database — plates render on playback with a jump-to-timestamp panel, store no new video, and stay inside the compliance line.
• Gave fleets three run modes — automatic, trigger-scoped, or on-demand: trading instant availability against processing cost.

Impact

Shipped to GA across 300+ enterprise fleets and 50k+ devices at 96% detection on clearly visible plates — cutting incident investigation time ~75%. Reviewing a claim went from minutes of frame-by-frame scrubbing to one overlay and a jump straight to the plate, and once fleets turned it on, ~98% kept it on. Built as a standalone detect-then-read pipeline rather than a player bolt-on, the same capability now extends to other video surfaces without a rebuild — a platform, not a one-off.

Reflection

The hard part was not the OCR(although that was tough too), it was making a plate extraction trustworthy under a strict compliance boundary. Turning that constraint into the product's definition (an overlay, not a plate reader) is what made it shippable, and building it as a reusable pipeline is what let one launch become something the rest of the video product plugs into.