Things that shipped, and one that didn't.

Four projects. Three are in testers' hands; the fourth was researched and dropped.

A—01

One of Them Is Lying

A bilingual detective game where the suspects are language models, and they lie.

On TestFlight

Problem: A suspect who lies has to lie the same way twice. The story has to hold across a whole interrogation and break in exactly one place.
Decision: Everything runs on the phone, on Apple’s ~3B Foundation Models. No server, no cloud model, no ad SDK. Universal conduct rules live in the engine; case files carry content only, and a linter fails the build if a rule leaks into one.
Outcome: 56 cases in English and 简体中文 — 112 files. A 24-archetype portrait library, four states each. Every case embeds a solve path the test suite replays against the live model.

A—02

A bilingual solid-food tracker two parents share, with nothing on anyone else’s server.

On TestFlight

Problem: Two parents, two phones, two languages, one picture of what a baby has eaten. Without an account, a subscription, or a company holding the record.
Decision: Local-first. iCloud does the syncing, each parent on their own Apple ID. No accounts, no backend, no subscription, no ads. Written against a spec before any code.
Outcome: 188 foods seeded, 55+ of them Chinese staples, 21 compound dishes. 91 unit tests. The screenshots here came out of an automated run that walks the logging flow in both languages.

A—03

Menu-bar dictation that runs entirely on the Mac.

Signed & notarised

Problem: Dictation is either fast or accurate. Streaming models give you words as you speak and cost precision; whole-utterance models are right and late.
Decision: Speech recognition sits behind one Swift protocol. The audio, hotkey, HUD and text-injection layers never learn which engine is running, and engines swap at runtime. Native Swift and SwiftUI on Core ML. No telemetry, no analytics, no network call except a model download you ask for.
Outcome: Signed, notarised, in testers’ hands. Two engines behind one interface. The same abstraction carried a C#/.NET port to Windows.

R—01

A trading strategy, researched and not deployed.

Not shipped

Problem: The same 17.5 months of market data had been searched about ten times. On the eleventh pass, a good-looking result is as likely to come from the searching as from the market.
Decision: The strategy was dropped and the failure mode written up as a standing rule sheet: discovery split from validation, and the arithmetic of multiple comparisons applied before any deploy decision.
Outcome: Nothing went live. What exists is a research loop that can tell a finding from a coincidence, and a written reason for every rejected candidate.

Send the problem. You will get a scope and a price.