Oleg Kossoy
Software Engineer · Self-Hosted Everything

I build real products on my own infrastructure — and I've taught an AI fleet to ship them.

Media pipelines with GPU workloads, native audio plugins, network tooling, a Windows/Linux media player — all designed, tracked, and deployed through Maestro, my self-hosted agent orchestrator that turns GitHub issues into reviewed, merged, deployed PRs. Nine products, one control room, minimal hands on keyboard.

9
live project flows under one orchestrator daemon
80
merged PRs in the last 7 days — mostly agent-authored
30+
hands-off releases shipped for a single product
5
Proxmox nodes running it all, at home
Featured Work

Products in production

Each of these runs today, on hardware I own, serving real users — starting with me.

AI Orchestration · Go

Maestro

The software factory behind everything else.

A single Go daemon that runs an entire fleet of software projects autonomously: it picks up labelled GitHub issues, spawns AI worker agents in isolated git worktrees, opens PRs, drives them through CI and an automated review gate, and merges — with a cautious human-approval gate for risky actions. Multi-backend model routing with automatic failover, SQLite config store, and a live Fleet Mission Control dashboard. It dogfoods its own development and self-deploys on merge.

Go · single binarySQLite 9 projects orchestratedreview-gated auto-merge
Maestro Fleet Mission Control dashboard
Fleet Mission Control — live supervisor heartbeat, approvals, per-project timelines
GPU Media Pipeline · Python + React

Karaoke

Any YouTube link → a stage-ready karaoke track in minutes.

Paste a URL and get isolated vocals, an instrumental playback track, and time-synced lyrics — BS-Roformer stem separation on serverless GPUs, LRCLIB lookup with forced alignment and Whisper fallback, and an in-browser performance console with live vocal fading and A–B loops. Runs to an SLO (p95 ≤ 6 minutes), costs cents per track, and has gone through ~30 releases with near-zero manual coding.

FastAPIReact SPA RunPod serverless GPUBS-Roformer + Whisper Web Audio
Karaoke console with split waveform, vocal faders, and synced lyrics
Console — split waveform, VOX/INST faders, hold-to-drop vocals, synced lyrics
Transcription Service · Python + React

Scribe

A private research library built from videos.

YouTube URL in; searchable transcript, LLM summary, and topic tags out. Whisper runs on spot GPU instances under a rolling spend cap; summaries flow through a model chain with automatic fallback. Full release automation — every merge tags a SemVer version, deploys, health-checks, and rolls back on failure. My personal watch-later queue is now a 280-transcript searchable knowledge base.

FastAPIWhisper on vast.ai spend-capped GPUauto-deploy + rollback
Scribe library of transcribed and summarized videos
Library view — transcripts, LLM summaries, tags, live pipeline spend
Network Operations · Rust + Next.js

Panoptikon

The all-seeing eye for the home network.

A self-hosted network operations console: ARP device discovery, MikroTik router management, lightweight host agents, on-demand port scans, DNS logs, QoS and NAT tables — all streaming live over WebSockets into a dark operator UI. Every feature PR ships with a mandatory Playwright end-to-end test; the repo contract enforces it.

Rust · axumNext.js 15 MikroTik APIWebSocket live updates Playwright E2E gate
Panoptikon live network dashboard
Live dashboard — 165 known devices, WAN traffic, top talkers, 12 ms WebSocket feed
Native Desktop · C# / WinUI 3 + Rust / GTK4

OK Player

An IINA-class media player for Windows and Linux.

Built on libmpv with two native shells — WinUI 3 with Fluent design on Windows 11, GTK4 on Linux — converging on a shared Rust core. Chapter thumbnails, frame-precise stepping, synced lyrics, A–B loops, mini-player, auto-updates via Velopack. Windows is at v0.10 with an external beta tester; the Linux shell has burned through 100+ alpha builds. Best-in-class subtitle UX is the hill it's built on.

libmpvWinUI 3 · Mica GTK4shared Rust core Velopack updates
OK Player playing The Apartment (1960) with chapter thumbnails and subtitles
Chapters with thumbnails, subtitles, bookmarks — The Apartment (1960)
Self-Hosted Gallery · Go + React

OK Folio

A beautiful folio for visual discoveries.

An image-first, self-hosted gallery for collecting and curating visual work from many sources through pluggable provider connectors. Go backend with Postgres and Valkey caching (pgvector groundwork for visual similarity), an editorial serif-led React UI with its own designed brand system, and a curation inbox that treats collecting as a deliberate act — 43,000+ pieces kept with intention.

GoPostgres + pgvector Valkeyprovider connectors custom brand system
OK Folio gallery with curated art pieces
Gallery — pieces kept with intention; magazine, library, and wall layouts
Audio Engineering & Tooling

Plugins, utilities, and sharp little tools

I play guitar and record music — so a lot of the tooling scratches real studio itches: DSP plugins in C++/JUCE, CoreAudio utilities, and rehearsal automation.

Apertune tuner panel — in tune, sharp, and flat states
Audio Plugin · C++ / JUCE

Apertune

A tuner that snaps into focus.

Studio-grade chromatic tuner (CLAP / VST3 / AU) with a rangefinder metaphor — pitch locks at a center reticle like a camera aperture. Custom tunings for 4–9 strings, 432–448 Hz reference, accessible non-color cues. Passes clap-validator 16/0.

JUCE 8CLAP · VST3 · AUDSP tests in CI
Audio Plugin · C++20 / JUCE private

S1A Delay

Hybrid analog/digital delay.

A professional delay plugin (VST3 / AU) with extensive DSP: modulation, analog-voiced emulation, tempo sync. C++20, CMake, JUCE 8 — validated with pluginval at maximum strictness and Steinberg's VST3 validator. macOS Apple Silicon first, Windows next.

JUCE 8C++20pluginval level 5
macOS Utility · Swift

KatanaDI

Your amp, straight to your Mac.

Menu-bar utility that routes the clean DI signal from a BOSS Katana over USB to the Mac's output via CoreAudio — no effects, no modeling, no DAW required. Auto-detects the amp, corrects clock drift, shows live buffer and peak diagnostics in the menu bar.

CoreAudiomenu-bar appuniversal binary
RehearseKit DAW-style stem mixer
Web App · Next.js + FastAPI

RehearseKit

From any song to a rehearsal-ready DAW project.

Upload audio or paste a YouTube URL; get 4-way stem separation (Demucs), BPM detection, waveform trimming, and a generated .dawproject that opens in Studio One, Bitwig, Reaper, or Cubase — with a DAW-style mixer for instant solo/mute preview in the browser.

DemucsCelery + Redis.dawproject export
Telegram Agent · Go

OK Gobot

Reliability over demos.

A Telegram-first AI agent as a single Go binary — a ground-up rewrite of a 197 MB TypeScript assistant into 18 MB that starts in 15 ms. Multi-model routing, durable jobs and cron, semantic memory, and self-evolving skills behind a safety-audit gate.

single binary~10 MB RAMmulti-model routing
Purchase Intelligence · FastAPI + React private

GearIQ

Know the real price before you buy.

A shopping companion for musicians importing gear from Thomann: builds a virtual cart with real-time predictions of total landed cost — shipping, customs, and fees — then tracks each order through delivery and learns from actual costs to sharpen the next prediction.

cost predictionorder lifecycle±10% target accuracy

How a feature ships here

GitHub issue Maestro spawns agent PR + CI automated review gate auto-merge deploy · health-check · rollback

Everything above runs on a five-node Proxmox cluster in my home — GPU bursts rented by the minute, secrets in Infisical, storage on ZFS, ingress through Cloudflare tunnels. See the lab →