Shipping facial recognition on $200 hardware
The model was never the problem. The per-device budget was.
- Published
- 13 July 2026draft
- Category
- AI in production
- Reading
- 9 min
This piece is published as an outline: the statement and the structure are settled, the full text is coming.
The budget rules
A customer with twelve sites will not buy twelve GPUs. The real constraint is not model accuracy: it is the cost of each capture point and how fast it answers with fifteen people queuing.
The pipeline that survived
MTCNN for detection, RetinaFace when the lighting is hostile. ArcFace embeddings. Matching against the site index, not the whole database.
Quantising without losing the face
What INT8 actually costs you, and why the similarity threshold has to be recalibrated afterwards, not before.
The latency budget
From frame to recorded punch. Where the time actually goes, which is rarely where you think.
Write to me
If you are thinking about building something with AI and want to talk it through with someone who has put it into production, write to me. I answer myself.
hola@giron.ai