01 / 07 · Personal · flagship · May – Nov 2026

SkyControl

Voice air-traffic control at SFO, running on small models I trained. No LLM, no GPU at run time.

The problem

Air-traffic control is a spoken job. Games turn it into menus.

You are ground, tower or approach. Key the mic, speak a clearance; pilots read it back and fly it.

I wanted that to work on an ordinary laptop, in a browser tab. So I trained my own small models for every step that listens, decides or speaks. The model only ever writes a typed command line like UAL214 C 5000. Everything else is rules, a game engine and a safety gate.

Simulation

Work the scope.

A simulated, in-browser demo of the pipeline. Rules-based stand-ins play the part of my models here; the real ones run at openatc.aeon-labs.site.

1 · Pick an aircraft

2 · Choose a clearance

Heard

—

Command line

—

Gate

idle

Pilot readback

Select an aircraft and transmit.

  1. Micpush to talk
  2. ATC-ASR 110M~0.2 s on CPU
  3. Clean-up rulesnumbers, fillers
  4. Command parserrules first
  5. S1-2Monly if rules fail
  6. Safety gatesend · confirm · say again
  7. Game engineflies it
  8. Phraseologyreadback text
  9. Paradee-8Mpilot voice, WASM

How it works

Three small models.
One typed line between them.

01 · Listen

ATC-ASR 110M

Speech-to-text tuned on air-traffic radio. 0.6% word error rate on real ATC recordings, about 0.2 s on a CPU. Clean-up rules normalise numbers and fillers before parsing.

02 · Decide

S1-2M, from scratch

A 2M-parameter GPT-style transformer trained on 205,000 simulated situations across 11 airports, with a custom 1,349-token BPE where every ATC word is one token. It runs int8 in a Web Worker in 0.1–0.2 s, and only when the rules can't parse.

03 · Check

A safety gate

Every command passes a verifier before the engine sees it: send, ask to confirm, or "say again". It cut wrong auto-sends from 19% to 0.9%.

04 · Speak

Paradee-8M voices

Phraseology turns the command into a proper readback, and an 8.5 MB TTS model speaks it at 6× real time in WASM, in the browser.

Numbers

Measured, not vibes.

0.6%

word error rate on real ATC recordings · 110M ASR · ~0.2 s on CPU

GRPO with the game engine as the reward · exact match

p = 0.013 against the SFT control

2M

parameters in S1-2M, trained from scratch

205,000

simulated situations across 11 airports

1,349

tokens in a custom BPE · every ATC word is one token

19% → 0.9%

wrong auto-sends after the safety / verifier gate

74 → 44

wrong-aircraft errors after GRPO

4.5× / 7×

faster decoding from a hand-written KV cache / prefix caching

8.5MB

Paradee-8M pilot voices · 6× real time in WASM

62ms

on-device streaming speech in the Flutter phone app

14 Nov 2026

first public release · MIT licence · models on Hugging Face

The reel

Thirty-two frames, one block of time.

Move over the stack to turn it; scroll to scrub through it. Each slice is a frame of the reel, stacked in depth.

Placeholder reel — swap for a gameplay capture

Stack

What it's made of.

  • TypeScript
  • Next.js
  • three.js
  • Mapbox GL
  • PyTorch
  • HF TRL / PEFT
  • ONNX Runtime Web
  • sherpa-onnx
  • Flutter
  • vLLM
  • Playwright

First public release on 14 Nov 2026, MIT licensed. Models are on Hugging Face.

Next project · 02 / 07 Stratify An MCP server for honest options backtesting, tested on real 1-minute prices.