Sergey Drozdov · @stereosam
Before this I ran businesses — sold STEREOSOCKS, a B2B brand; still run MIXY. I didn't come to this from a dev job: the first thing I automated was my own operations. That is also why I killed my own feature once the unit economics stopped working — see below.
Now I run several coding agents at once, on a harness I built for it. Five months of vibe coding, 34 projects, 127 tasks closed in August across 10 of them — all of it under one meta-brand, STEREO AI. Products, tools and the workshop live under it; the repositories below are parts of the same thing, not side projects.
How I work. One master memory store, every machine a read-only mirror — one
writer, so conflicts can't happen by design. Each project gets three files with
three different lifespans: status (current truth, overwritten), backlog (work),
history (append-only). An idempotent engine writes back. Eleven slash commands of
my own. A Telegram bridge to the Agent SDK, so the same verbs work from my phone.
The trick. Every mistake I make becomes a written rule the agent reads before it works again. 57 so far, sharded by topic, loaded only when relevant. A mass mailing once reached the wrong contacts — the fix was a rule, not a patch. It hasn't happened since.
The harness itself
| stereo-vibe-public | One-click starter kit: sets up an AI-coding environment on Windows/macOS and scaffolds the agent's memory from what it just detected about the machine |
| stereo-reels-public | Agent skill: long recording → short vertical clips, where every cut is transcribed again to prove it isn't broken |
Products, shipped to people who pay
| stereo-dictator-public | Voice input and transcription. Tauri 2 + Rust: hold-to-talk into any window, batch files with speaker diarization. Ships with a mock server, so it runs without a GPU backend. Swapping the ASR engine took a 12-minute file from 58.3s to 29.3s, VRAM 3.6 GB → 0.49 |
| stereo-sync-public | Two-way folder sync between cloud and desktop. Tauri 2 GUI over rclone bisync, OAuth PKCE, signed auto-updates. Written the week a vendor removed its free desktop client |
The feature I killed
I A/B'd a cloud speech recogniser against my own GPU, liked the quality — and rolled it back. Per-minute minimums meant short dictations would cost about 14× more. Months of my own work went in the bin because the unit economics did not survive contact with how people actually use the thing. Coming from running a business, that call is reflexive; it is the same instinct that makes retention bugs feel urgent.
Not on GitHub: STEREO Practicum — about 50 people taught in person to work this way.
stereosam.github.io — who I am, in one screen.
Novosibirsk, UTC+7. Python, Rust/Tauri, PowerShell, nginx, systemd, CUDA.