Product build

WatchPost
A doorbell that remembers who came by

Solo build Python YOLOv8 GPT-4 Vision 2025

A doorbell camera watches everyone who walks up, then forgets them as soon as the clip ages off the feed. WatchPost sat on top of that feed and kept what the camera threw away. It spotted the person, looked at what was on screen, and wrote down what it saw.

Architecture: the doorbell RTSP feed runs through YOLOv8 person detection, a confidence gate with a cooldown, GPT-4 Vision reading the still, face clustering, into a SQLite store with an image gallery, browsable as a log by day.
How it worked: detect, read, recognize, log, browse by day.

The problem

Finding one visit means scrubbing through hours of empty driveway. The camera records all of it but lets you search none of it, so the clip you actually want is the one you can never find.

The build

The feed came in over the camera's video stream. A detector watched for a person in frame and only woke up when it was reasonably sure one was there, with a short cooldown so a single visitor did not trigger it fifty times. When it fired, GPT-4 Vision read the still: what the person was doing, what they were wearing, whether anything looked off.

Faces got clustered, so the same person showing up twice was recognized as the same person rather than two strangers. Every encounter landed in a small database with its image, and a plain web page let you browse the whole thing by day.

The outcome

Instead of hours of footage, you got a log: who came by, when, and what the camera saw. The detection was straightforward. The work that mattered was the writing-down, turning a feed that used to forget everything into a record you could read back.