iOS tech lead · building with AI agents
I lead iOS development at a bank and build products with coding agents that work from specs and tests. This is where I write down what works, what didn't, and why.
A dinosaur learning game for children aged 4 to 7.
React · TypeScript
An offline iPhone app that reads your notes aloud.
SwiftUI · Swift Testing
$ agister run review
resolve ✓ 3 definitions
admit ✓ budget 1.20 USD
record runs/7f2c/
A runner for AI-agent work defined in settings files.
Go
Learning programs your own agent can work on.
Node.js · SQLite
An iPad and an iPhone agree a payment over Bluetooth.
SwiftUI · Core Bluetooth
Agonara Session lets a student read an Agonara course on a phone, first in a Telegram bot, then as a web page on the local network. Why a lecture is paged by screen rather than by section, what Telegram messages can and cannot do, and how a Go runner redeploys itself when trunk moves.
Audelin's live transcript used to rebuild the whole call whenever a word changed. Making the work follow the live tail brought a late view update from hundreds of milliseconds to about one.
A phone that pays at a till must never pay the till next door. How Mizan BLE picks the nearest host by smoothed signal strength, what a real iPad and iPhone measured, why the first picker flickered at arm's length, and what iOS does not let a Bluetooth payment do.
A payment over Bluetooth has one way to succeed and many ways to stop. How Mizan BLE splits each side into a chain of state machines, why the countdown stops once the amount is sent, and how an iPad ends a link it is not allowed to close.
Mizan BLE is a proof of concept in which an iPad at the till and a payer's iPhone agree a payment over Bluetooth Low Energy. What it is, which device plays which Bluetooth role, and how the four characteristics of one service carry a whole transaction.
Agonara can grade a student's open answer with any model behind an OpenAI-compatible endpoint. The model grades per criterion, its output is validated before a student sees it, and the work runs as a leased, retried, budgeted job.
In Agonara, model suggestions and curator drafts never become program content by themselves, accepted content is kept as revisions, and a student's answers are pinned to the revision they answered. How the modules keep their own data apart.
Agonara is a bring-your-own-agent environment for learning programs. A curator builds a program from sources, students follow it and prove they understand, and the agent you already use does the drafting, quizzing and grading. What it is and why only a person accepts content.