BUILD RECORD
WatchPost
A camera produces hours of footage but no usable record. This first pass turned visits into searchable events.
SUMMARY
WatchPost
Retired · built solo
Nothing expensive ran until a person appeared.
A vision layer on a live doorbell feed. Each visit became a record you could look up: the still, what the person was doing, wearing, anything off.
Proved the premise, and its own architecture wrong. Sampling a live stream made a worse record than the video already on the camera, so Doorbell Intelligence read the originals instead.
One feed, my own porch. One entry per visitor, not per frame.
A cheap detector on the stream, an expensive model on the still. YOLOv8, GPT-4 Vision, face clustering.
Solo. Architecture, detection, vision calls, storage, interface.
Ran live on my own porch feed. Every encounter stored with its image, browsable by day.
THE METHOD
A camera feed became data
a camera feed · the path to data
watched · gated · described · stored
Took the live feed
YOLOv8 watched the doorbell stream for a person.
Found the part carrying data
A confidence threshold and a cooldown let one visitor fire once.
Read it into words
GPT-4 Vision described the frame; clustering grouped repeat visitors.
Wrote it down
It wrote encounter, image, and date into a small database.
LIMITS
What it refused to do
- Clusters carried no names. It identified nobody.
- Below the detection threshold, nothing was logged. It missed visits before it invented them.
- The log came from sampled frames, not the camera's own recording.
- My own porch feed. Stills and log stayed private.
RECORD AND LINEAGE
The idea survived the build
Doorbell Intelligence, the same premise against the camera's original recordings.
FrameIQ and vLLM Video Intelligence, at channel scale.
Retired. Both are off now.
THE ARCHITECTURE
A RELATED PROBLEM
Need to find an event without reviewing the entire archive?
This build is retired, but the archive problem is not. If video needs to become evidence people can retrieve, I would like to hear about it.