Delos — DIY Voice-Note Device
A handheld voice-note device I specced, assembled and debugged — a Waveshare ESP32-S3 board with a 1.54" e-paper screen in a 3D-printed case. I specced it, built it and debugged it on real hardware; Claude wrote the firmware and the server side with me directing it. Hold record, speak, and the audio goes over WiFi to a FastAPI service on the homelab, gets transcribed by a self-hosted Whisper container, cleaned up and titled by a single Claude call, and written into my Obsidian vault. Double-tap to ask a question and the answer renders on the e-paper. Firmware updates itself over the air with bootloader-enforced rollback.
[ hold REC ] ─▶ 16kHz mono wav ─▶ WiFi
│
voice-notes (FastAPI, token auth)
│
Whisper (faster-whisper, int8, local)
│ transcript
claude -p ─┬─ title + cleanup
├─ tags + [[wikilinks]]
└─ tasks + reminders
│
Obsidian vault ─▶ laptop + phone (LiveSync)
sleep screen ◀── server-rendered panel (agenda · tasks · battery)
new firmware ─▶ device pulls on next sync ─▶ rollback if it won't boot This is the project I didn't expect to build. I wanted something I could talk to without picking up my phone, because picking up my phone means twenty minutes gone. So Delos is deliberately limited: it captures, and it never hands you anything back to scroll.
How it got made: Claude wrote the code, I directed it. Mine are the idea, the hardware, every call about what this should and shouldn't be, and the part where it's on the desk not working and someone has to figure out why. I've killed finished features for making it too much like a phone — that filter is the project.
Hardware is a Waveshare ESP32-S3 with a 1.54" e-paper display in a case I printed. Hold the button, talk, let go. The clip goes to a FastAPI container on the lab, Whisper transcribes it locally — no cloud speech service — and one Claude call cleans up the rambling, titles it, tags it, links it to related notes and pulls out any to-dos. It's in my vault and on my laptop before the device is back in my pocket.
The screen isn't wasted while it sleeps. The server renders an ambient panel — today's calendar, open tasks, battery — and the device just draws whatever it's handed. That was the rule the whole way through: dumb device, smart server. Anything I want to change about what it shows is a Python edit and a container restart, not a reflash.
It also updates itself. I publish new firmware, the device notices on its next sync, flashes it, and the bootloader rolls it back on its own if the new image won't boot — which has already saved me. Double-tapping record asks Claude a question out loud and the answer shows up on the e-paper in about twelve seconds.
Most of the lessons were budgets I didn't know I had. E-paper is free to hold an image but the radio isn't, so every sync is battery. And a 200-pixel-wide screen will quietly eat the end of any line you didn't actually measure in the font you're rendering with.