SFO · ground · tower · approach

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.

Role
Everything — models to UI
When
May – Nov 2026
Released
14 Nov 2026 · MIT
N 090 180 270
“United two fourteen, climb and maintain five thousand.” UAL214 C 5000
01

The problem

Filed as flight progress strips — the paper ATC still trusts.

  1. ATC01B738/M
    REQVOICE

    Air-traffic control is spoken.

    Every clearance is a radio call. A game about it should listen, not offer buttons.

    KSFO
  2. LLM02H/GPU
    REQDENY

    A big model is the wrong tool.

    A general LLM needs a server and a GPU, and it can invent an aircraft. I wanted every turn to run in a browser tab.

    NO GPU
  3. SML032M/L
    CLRSHIP

    So I trained the small ones.

    A 110M speech model, a 2M command model and an 8.5 MB pilot voice. The model only ever writes a typed command line, like UAL214 C 5000.

    ON DEVICE
02

How a clearance flies

Mic to pilot voice, all in the browser except speech-to-text.

01 · Mic
  1. 01Micpush to talk
  2. 02ATC-ASR 110M0.6% WER · ~0.2 s CPU
  3. 03Clean-up rulesnumbers, callsigns
  4. 04Command parserrules first
  5. 05S1-2Min browser, if rules fail
  6. 06Safety gatesend / confirm / say again
  7. 07Game engineaircraft obey
  8. 08Phraseologyreadback text
  9. 09Paradee-8Mpilot voice, WASM
03

A 2M-parameter controller’s ear

S1-2M is a GPT-style transformer I trained from scratch on 205,000 simulated situations across 11 airports. It runs int8 in a Web Worker and answers in 0.1–0.2 s.

I wrote a custom 1,349-token BPE so every ATC word is one token. A hand-written KV cache made decoding 4.5× faster; prefix caching made it 7×. Then GRPO, with the game engine itself as the reward.

s1-2m.worker · int8 · onnx

tokenizer 1,349 tokens · 1 ATC word = 1 token

train 205,000 situations · 11 airports

decode KV cache 4.5× · prefix cache 7×


GRPO · reward = game engine

SFT control82.0%
before GRPO79.6%
after GRPO83.8%

exact match · p = 0.013 vs SFT control

04

Instruments

Measured, not estimated.

Word error rate · real ATC recordings

0.6%

110M ASR · ~0.2 s on CPU

Wrong auto-sends

19%0.9%

after the safety / verifier gate

Wrong-aircraft errors

7444

with GRPO on the game engine

Command model

2M params

0.1–0.2 s · int8 · Web Worker

Pilot voices · Paradee-8M

6× real time

8.5 MB · WASM

Phone app · on-device speech

62ms

Flutter · streaming

Custom BPE vocabulary

1,349tokens

every ATC word is one token

05

On frequency

Placeholder reel. The real one shows a full approach session, mic to touchdown.
07

Stack

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

Next application · 02 / 07

Stratify
stratify.mcp

An MCP server for honest options backtesting.