Project 01 · Personal · May – Nov 2026
SkyControl
Voice air-traffic control at SFO, running on small models I trained. No LLM, no GPU at run time.
You are ground, tower or approach. Key the mic, speak a clearance; pilots read it back and fly it.

- Speech
- ATC-ASR 110M
- Command model
- S1-2M · 2M params
- Pilot voice
- Paradee-8M
- Run time
- CPU + browser · no GPU
- Licence
- MIT
- Role
- Everything, end to end
- YOUUnited two one four, climb and maintain five thousand.
- S1-2MUAL214 C 5000
- UAL214Climb and maintain five thousand, United two one four.
Notes
An ATC game that actually listens.
- 1
Radio is clipped, fast and full of callsigns. I wanted a game where you talk like a controller and the aircraft simply do what you said.
- 2
A large language model in the loop is slow, needs a GPU somewhere, and can pick the wrong aircraft. That last one is the error that matters.
- 3
So I trained small, specialised models and put rules and a safety gate around them. The model only ever writes a typed command line.
Plan view · how it works
Mic to readback, fix by fix.
- MIC
Key the mic
Push to talk. You are ground, tower or approach at SFO.
- ASR
ATC-ASR 110M
Speech to text, trained for radio. 0.6% WER on real ATC recordings, about 0.2 s on CPU.
- RULES
Clean-up rules
Spoken numbers, callsigns and phraseology are normalised before anything else runs.
- PARSE
Command parser
A deterministic grammar goes first. The model only steps in when it fails.
- S1-2M
Command model, in the browser
A 2M-param transformer I trained from scratch writes one command line. int8 in a Web Worker, 0.1–0.2 s.
- GATE
Safety gate
Send, confirm, or “say again”. Wrong auto-sends fell from 19% to 0.9%.
- ENGINE
Game engine
The aircraft flies the clearance. The same engine was the reward during RL.
- PHRASE
Phraseology
The typed command is turned back into a proper readback.
- VOICE
Paradee-8M pilot voice
8.5 MB of TTS at 6× real time in WASM. The pilot reads it back, in the browser.
Detail A · the only thing the model writes
One word, one token. One line out.
I built a custom 1,349-token BPE where every ATC word is a single token. The model reads the situation and the transcript, then writes a typed command. Nothing else can reach the game.
Transcript in
- united
- two
- one
- four
- climb
- and
- maintain
- five
- thousand
Command out
UAL214 C 5000
Profile view · what runs where
Every model runs on your machine.
Minimums · measured results
Small models, measured.
| Component | Value | Conditions |
|---|---|---|
| Speech · ATC-ASR 110M | 0.6% WER | Real ATC recordings · ~0.2 s on CPU |
| Command · S1-2M | 2M params | GPT-style, trained from scratch on 205,000 simulated situations across 11 airports |
| Tokeniser | 1,349 tokens | Custom BPE · every ATC word is one token |
| Decoding | 4.5× · 7× | Hand-written KV cache · prefix caching |
| RL · GRPO | 79.6 → 83.8% | Exact match, game engine as reward · SFT control 82.0%, p = 0.013 |
| Wrong aircraft | 74 → 44 | Errors, before and after GRPO |
| Safety gate | 19% → 0.9% | Wrong auto-sends |
| Pilot voice · Paradee-8M | 8.5 MB | 6× real time in WASM |
| Phone app · Flutter | 62 ms | On-device streaming speech |
Axis from 75%. p = 0.013 vs SFT control.
Tower view · reel
Cleared to land.
Ramp · screens
Ground, tower, approach.






Communications · stack
Built with
- Web
- TypeScript · Next.js · three.js · Mapbox GL
- Training
- PyTorch · HF TRL / PEFT · vLLM
- Inference
- ONNX Runtime Web · sherpa-onnx
- Mobile
- Flutter
- Testing
- Playwright
Released 14 Nov 2026 under MIT. Models on Hugging Face.
