Door camera
A camera at the front door that answers who is there rather than just something moved — and answers it on a small computer in the house, because the alternative is streaming my family and everyone who visits to somebody else's servers.
- Role
- Sole author
- Status
- Working, paused
- Inference
- On-premise
- Built
- 2026
What it does
A small camera at the door streams video to a slightly larger computer inside the house, which runs person detection on a dedicated AI accelerator and then recognizes who it is looking at. The result becomes a notification through my own home automation system: not "motion detected," but a name.
Nothing leaves the property. There is no subscription, no vendor holding footage of my front step, and no account that can be breached somewhere else.
The decision worth explaining
This project handles face data of real people — family, friends, whoever comes to the door. That is biometric information, the category you cannot reissue after a leak. You can change a password; nobody gets a new face.
Fail closed. If the access token is not configured, the endpoint that lists known people does not fall back to open — it refuses. An unconfigured system is a locked system, not a public one.
Biometric data never enters the repository. The enrollment photos were moved out of the project tree entirely, and the recognition database is excluded from version control.
The default direction of a failure is a design decision, and it is one people get backwards constantly — a missing configuration value quietly disabling a check is how systems end up unprotected without anyone choosing it. I also went back and added authentication and cross-site-request protection to the control endpoints after the fact, plus a regression test, so the protection cannot silently disappear in a later change.
Why it is paused, not abandoned
It works. It sits at a natural stopping point: the hardening pass is done, the runbook is written, and the next phase — moving detection onto the camera sensor itself to save bandwidth and power — is an optimization rather than a fix. I would rather leave something documented at a clean boundary than half-migrated.
Honest limits
- Recognition accuracy depends heavily on light and angle. A backlit face at dusk is a coin flip.
- It is a single-household system on a trusted network segment, not a product.
- Enrollment is manual and deliberately awkward — adding a person is a decision, not a convenience.
- The detection hardware is a specific accelerator board; it is not portable to arbitrary machines without rework.
Seeing the code
The repository is private. If you are evaluating me and want to read the code or walk through the architecture, ask me and I will arrange read access or a screen share.