CCTV IQ - Face ID Attendance

At a glance

In short	A face-recognition attendance system on two office doorway cameras. It logs clock-in, clock-out and time inside, and answers questions about attendance in plain English.
Problem	Track office attendance automatically, using the cameras already at the door.
What it does	Recognises enrolled people on two doorway cameras, records clock-in, clock-out and time inside with a face photo for each, and writes one Excel file per day.
How it works	Find faces, turn each into an embedding, match it against 182 people. Clock-in is the first sighting of the day on either camera; clock-out is the last.
Ask Iris	Type a question like "who was late this week?". Claude Haiku 4.5 on AWS Bedrock writes one SQL query, runs it, and answers from the rows.
Result	0.31 s per frame (from 9.86 s) on an office PC. A score floor fitted to a hand-checked day removed every false name and lost no real person.
Tech	Python, InsightFace (SCRFD + antelopev2), OpenVINO on the Intel iGPU, FastAPI, SQLite, AWS Bedrock

Works with either door both ways: the first and last sighting on any camera decide the day
Time inside = first to last seen, minus the breaks the cameras saw; a missed exit never docks an honest employee
Masks the clear glass strip where the lobby camera could see people at their desks, which had cut about 1 in 8 long days short by 2.7 hours
Refuses a name when two people look too alike: no name beats a confident wrong one
Runs unattended: starts at login, a supervisor restarts it in about 15 s, and the dashboard never says it is recording when it is not
Iris never sees staff data to decide what to fetch, only the table layout; it declines off-topic or destructive requests

Speed: where the models run

Setup	Seconds per frame
Both models on CPU	9.86
Recogniser on iGPU, detector on CPU	0.69
Both on iGPU	0.31

Picking the score floor from a hand-checked day

Floor	Names kept	Wrong names kept
0.24	70	26
0.35	55	11
0.40	47	2
0.41 (chosen)	45	0

Accuracy: face size matters most (17,842 real sightings)

Technical details

The older in/out pairing fully measured only 14% of person-days; the new span-minus-breaks rule gives every day a number and agrees with pairing within 15 minutes 94% of the time
A break needs a lobby sighting, then 10+ minutes unseen, then proof of coming back in; days that could hide an early exit get a check flag instead
Switching both cameras from H.265 to H.264 fixed 60-73% of frames arriving with a broken bottom band (ping and substream tests had ruled out WiFi)
Found and fixed a 9-day silent outage: a thread that crashed quietly, a status flag set before the work started, and nothing watching the process
A stalled camera is rebooted over HTTP, and reconnects back off instead of hammering it
Iris retries a failed query with the database's own error; every checkable answer in the tests has hand-written SQL beside it
Compared four recognition models (antelopev2, AdaFace, buffalo_l, AuraFace): all tied within noise, so camera placement matters more than the model
Real phone photos matched about 9.5 points better than AI-edited directory photos
Recognition runs in its own thread; the dashboard only reads snapshots, so closing the browser changes nothing
Regression tests cover the visit logic, the arrival rules and the ignore-zone geometry

The code for this project is private, so there is no public link.