A security pipeline that recognizes, reviews, and learns.

Cameras detect objects, a recognition layer matches faces against an enrolled identity memory, a GenAI model narrates each scene, and anything uncertain routes to a human for a one-tap decision — which feeds active learning. Event ingestion, classification, human review, and an audit trail: the same shape as financial controls and exception monitoring.

● Synthetic walkthrough — no real footage, identities, or locations. All events and faces are fabricated.

Part 1Event → recognition → alert

A camera detects an object (Frigate), the event flows through Home Assistant into Harley Vision, a face is matched against enrolled identities (CompreFace), a GenAI model describes the scene, and a Telegram alert arrives with one-tap review buttons.

Part 2Console & active learning

Every event lands in an operator console — the system of record. Known faces are remembered; unknowns go to a review queue. Labeling an unknown once enrolls it as a new example, so recognition improves on the next sighting.

§ 02 · the screens

The live eventFrigate classifies the object and draws a detection box; the event flows Frigate → Home Assistant → Harley Vision, where CompreFace matches the face and a GenAI model writes a plain-language scene description.

Telegram reviewThe alert reaches a phone with the recognition result and one-tap buttons — Expected / Flag / Label — so a human stays in the loop on anything uncertain.

Operator consoleA review queue, an identity memory store (enrolled household members and recurring visitors), and daily stats — a proper system of record for events, identities, and reviews.

The stackPython · FastAPI + uvicorn · PostgreSQL (SQLAlchemy) · Frigate object detection · Home Assistant · CompreFace face recognition · GenAI scene descriptions · Telegram review bot · systemd. Ingest → classify → human review → audit trail is the same shape as financial controls and exception monitoring.