Uploaded June 2026 | Updated September 2026, 2 weeks ago
Point a few phone cameras around a room and you've got a working AI security setup, no extra hardware.
Each camera sends its stream over WHIP into a Smelter compositing engine on the server, Smelter merges them into one output you watch live over WHEP, and a bit of server-side Python scores motion in real time. Cross the threshold and the focus jumps to wherever the movement is, the recorder kicks in, and the clip lands as an MP4. Then Gemini 2.5 Flash watches it back, writes up what happened, and assigns a severity rating. If it's serious enough, a Web Push notification fires even with the tab closed. Recordings, analyses, settings, camera names, all persisted in SQLite.
The fun part: it's the same engine, the same pipeline, the same handful of files. We dropped a few music videos into the local inputs folder, tweaked one line of the Gemini prompt to flag the most chaotic moments, and the whole thing happily started scoring dance routines and pyrotechnics instead of someone walking through the kitchen. Swap the inputs, change one prompt, and the pipeline analyzes something completely different from what it was built for.
So if you ever need your own monitoring setup, multiple cameras, automatic recording, AI event analysis, with no off-the-shelf NVR software, this is roughly plug and play. Or without the plug.
The app was inspired by Patryk's idea at a Software Mansion x Gemini hackathon earlier this year. As of the latest Smelter release, which landed a few days ago, this kind of pipeline is supported natively, and the next videos go into more complex setups.
Part of the Smelter Workshop series.
Links:
🔗 Smelter: https://smelter.dev
🔗 Live Demo: https://workshop.smelter.dev
Repos:
🔗 Project repo: github.com/M0rn1ngStur/smelter-surveilance
Point a few phone cameras around a room and you've got a working AI security setup, no extra hardware.
Each camera sends its stream over WHIP into a Smelter compositing engine on the server, Smelter merges them into one output you watch live over WHEP, and a bit of server-side Python scores motion in real time. Cross the threshold and the focus jumps to wherever the movement is, the recorder kicks in, and the clip lands as an MP4. Then Gemini 2.5 Flash watches it back, writes up what happened, and assigns a severity rating. If it's serious enough, a Web Push notification fires even with the tab closed. Recordings, analyses, settings, camera names, all persisted in SQLite.
The fun part: it's the same engine, the same pipeline, the same handful of files. We dropped a few music videos into the local inputs folder, tweaked one line of the Gemini prompt to flag the most chaotic moments, and the whole thing happily started scoring dance routines and pyrotechnics instead of someone walking through the kitchen. Swap the inputs, change one prompt, and the pipeline analyzes something completely different from what it was built for.
So if you ever need your own monitoring setup, multiple cameras, automatic recording, AI event analysis, with no off-the-shelf NVR software, this is roughly plug and play. Or without the plug.
The app was inspired by Patryk's idea at a Software Mansion x Gemini hackathon earlier this year. As of the latest Smelter release, which landed a few days ago, this kind of pipeline is supported natively, and the next videos go into more complex setups.
Part of the Smelter Workshop series.
Links:
🔗 Smelter: https://smelter.dev
🔗 Live Demo: https://workshop.smelter.dev
Repos:
🔗 Project repo: github.com/M0rn1ngStur/smelter-surveilance
