TECHNICAL GUIDE
Edge AI Video Analytics Deployment & PoC Checklist
For technical buyers preparing an edge video analytics pilot and acceptance plan.
1. Fix the business question and evidence rights
Name the event, site owner and acceptance owner. Define positives, negatives, ambiguous cases and the operator response. Confirm permission to process the video and a separate right to publish any demonstration. A pilot does not authorize public use of a customer’s footage, plates or logo.
2. Freeze inputs and the runtime
Record board revision, OS image, drivers, decoder path, codecs, resolution, frame rate, software build and model hashes. Preserve the exact authorized test input. Build success, model initialization, actual inference and business acceptance are distinct gates and should have separate results.
3. Review model and package licensing
List YOLOv8, HyperLPR3, RKNN dependencies and any specialized assets used by the selected path. Check source license, weight provenance, conversion/distribution terms and required notices. Keep unresolved assets outside the agreed deliverable; changing the displayed model name does not solve licensing.
4. Validate rules and event lifecycle
Annotate ROIs, thresholds, entry/exit logic and missing-observation behavior. Include repeated detections, two vehicles in sequence, temporary occlusion and genuine departure. Review event splits, duplicates, misses and evidence association, not only whether a box appeared.
5. Check interfaces and recovery
Exercise analysis start/stop, camera loss, reconnect, process restart, receiver errors and upgrade/rollback where included. Agree JSON fields, authentication, evidence access and receiver deduplication. Offline buffering and retry guarantees must be explicitly implemented and tested, not inferred from local inference.
6. Measure performance honestly
Write down duration, concurrency, codec/model combination, resource measurements and the definition of latency. Separate an alert-time target from a measured distribution. A smoke run is not a thermal soak or long-term reliability test. Do not extrapolate a single stream or one easy clip to a site-wide capacity promise.
7. Close with a decision record
Deliver a passed/failed/untested matrix, known limitations, required assets, license decisions and a revalidation plan. Discuss these results with Hank before expanding scope. A useful pilot can identify a blocker; it should not hide failures to produce a favorable demo.
Use a decision table at handover
For each acceptance row, record the input identifier, package version, expected event, observed event, evidence reference, result and unresolved issue. Mark a case untested when the required hardware, license or authorized sample is missing. Keep the decision owner and retest condition clear. The pilot result should support a go, revise or stop decision, with no pressure to turn a resource check into a runtime pass or a short successful sequence into a reliability claim.
DEFINE ONE TESTABLE WORKFLOW
Bring your application to Hank.
Agree inputs, events and acceptance before discussing integration.