I let AI build a tool to help me figure out what was waking me up at night

You wake up at 3am with no idea what woke you. By the time your brain catches up, the noise is gone. So I let AI help me build a tool to finally answer that question, and to act on the answer. I try to pay attention to the small things that affect my quality of life. When something keeps bothering me, I want to investigate, find a likely cause, and act on it. What changed recently is what I'm willing to build to support that. With AI tooling, projects I would have dismissed a few years ago as "too much effort for the payoff" now fit into a weekend. So whenever I bump into a problem in my daily life, I catch myself thinking, "actually, I could build something to look into this". This post is about one of those problems: my sleep. I live in a noisy city. Some nights I wake up at 3 am with no idea what woke me up. Other nights I don't fully wake up, but my watch the next morning shows that something pulled me out of deep sleep at 3:32 am. And without naming the cause, you can't fix it. Was it inside the flat? Outside? A neighbor? A truck? A door? Any "solution" you try is just a guess, and guessing tends to be expensive. One bit of context first: I already have a smart home setup with Home Assistant and a bunch of sensors around the flat (motion, doors, lights, temperature, humidity, CO₂, air quality). A lot of the data I needed for this project was already being collected. I only had to add the audio piece, pull in my sleep data, and tie everything together. When the Pi hears something loud enough to be interesting, it saves a short clip with a few seconds of context before and after. The whole detection mode is gated by Home Assistant: the Pi exposes itself as a Home Assistant integration, and an automation only switches it on when I'm at home, in bed, and around my usual sleep time. Outside of those conditions, it's completely disabled, no microphone access at all. That's the behavior I wanted, even on my own home network. The web app is where the actual usefulness happens. Each night is laid out like tracks in a music editor: one for sleep stages, one for heart rate and HRV, a few for the sensor events, and one for the noise events with the audio loaded in. The most useful feature, by far, is how it visually marks the moments where my sleep stage shifted or I woke up. That's the entry point. I scan a night, find the highlighted moments (in red), and click to listen. One more nice touch: the frontend is a progressive web app with web push, served only inside my home network. The moment I wake up and check my phone, I get a notification that last night's data is ready to review. Nothing leaves the network, which I really like. To set expectations clearly: AI is what made it realistic for me to build this in a weekend. The whole thing took roughly 8 hours of work, plus a few small improvements over the following days. Without AI tooling, I would not have started. But I'm not using AI to identify the actual sounds (at least, not yet). The listening part, recognizing a door vs. dishes vs. a motorbike, is still me with headphones on. The tool just points me at the moments worth listening to. And on the workflow: I didn't read the code (conscious choice). I tested the results, gave direct feedback when something was off, and let the AI verify its own output by letting it take screenshots of the running app in my browser. For the micro-computer side, I went one step further. The Raspberry Pi was a fresh, empty install, so I gave the coding agent SSH access and let it test things directly on the device. It would set up an experiment, ask me to shout, drop something, or run the kitchen tap, record the sample, and then analyze it for me, sometimes pulling out spectrograms. I had to be explicit about wanting it to work this way, but once instructed, the iteration loop was honestly quite fun. For a personal project running on my own hardware, this whole setup was enough. The interesting shift here isn't that AI solved my problem. It's that AI lowered the cost of building the thing that lets me solve my own problem. I get the sleep data from my Garmin* watch. Every watch and ring calculates sleep slightly differently, and to be honest, I don't fully trust any of them on the exact sleep stage I was in at any given second. What they're all reasonably good at, though, is detecting when you actually woke up. Those wake events, plus the rough transitions between stages, are what I actually care about here. They're not a clinical truth. They're visual markers that tell me "this moment is worth investigating". Without them, I'd be sitting through hours of audio of the fridge humming and a neighbor doing nothing in particular. I'm not trying to do sleep science. I'm trying to find what sometimes made me feel rough in the morning. Once I started actually using this, the patterns came out fast. The usual suspects: With actual data, I could finally act with some confidence instead of guessing: It's not perfec

I let AI build a tool to help me figure out what was waking me up at night

Key takeaways

Also covered by

I let AI build a tool to help me figure out what was waking me up at night

Key takeaways

Also covered by

More in computer-science