A coding agent that lives on your computer — and can actually run it.
Matched a comparable open agent on SWE-bench Verified using 5.3× fewer tokens — ~110K against ~588K per instance, metered at the HTTP boundary by a proxy I wrote. Local: no account, no telemetry, zero third-party dependencies in the core.
What using it looks likeplain English in, verified work out
Its own coding agent built the desktop console, the browser control, the screen recorder and the iOS app. The feature list is the benchmark.
Dictation for developers. Hold a key anywhere in the OS, speak, and the text arrives at your caret. The speech model runs locally, so no audio leaves the machine.
Sold and shipping at 0.4.9. Three Rust crates, one signed native binary per OS, and the whole checkout-to-license path — Stripe through Cloudflare Worker through Keygen — wired end to end.
Push-to-talk, running herelocal speech engine in the shipped app
An AI buyer's agent for US homebuyers. A two-minute onboarding chat, then 24 matched homes, each with a transparent valuation and walk / climate / school signals.
Six domain agents fuse 60+ public-data sources — FEMA flood, USGS seismic, Cal Fire, OpenAQ, school attendance boundaries — behind a gradient-boosted valuation with a leakage-safe temporal split, served in about 10 ms.
An open-source, crisis-aware mental-health AI. 988 and Crisis Text Line escalation, conversations encrypted on the device, free and public-good.
Every safety rule is bound to a behavioral test in CI — the policy is something that runs, not something written down. The app gates on age and risk before it will talk to you at all.
Safety rules bound to CIthe policy runs
RLdota-rl
PyTorchOffline RLResearch
A Dota 2 bot trained offline from high-MMR human replays — behavioral cloning into offline RL, the AlphaStar-Unplugged pattern.
Self-play is throughput-bound because there is no fast simulator, so the real engine is used only for validation. Sanctioned path throughout: Lua Bot API and a dedicated server, no injection, no input simulation, no memory reading.
Offline, not self-playwhy the engine is only a validator
—replay frames
→
BCbehavioral cloning
→
Offline RLGPU-bound
→
Enginevalidation only
Sanctioned path throughout — Lua Bot API and a dedicated server. No injection, no input
simulation, no memory reading.
▶Video pipeline
PythonCUDAsystemd
Whisper → Claude → a text-to-speech cascade that degrades through three engines instead of failing a batch → FFmpeg → upload.
1,321 videos published unattended across 23 accounts, VRAM-coordinated against concurrent GPU work on a single workstation running about 50 systemd units.
Publication matrix23 accounts × 18 channels
— videos published unattended.
CKCartek
LaravelAWSStripe Connect
The auto marketplace I co-founded in 2018 and ran engineering for until 2025 — marketplace, ML, payments and infrastructure from zero.
Six engineers, 5M+ read requests a day at 99.9% uptime, on a double-entry ledger with Stripe Connect as merchant of record.
