Writing

Babbla's fast cleanup model was English-only and failed silently on Swedish: it handed the transcript back untouched, and an untouched transcript passes every guard. So I fine-tuned it bilingual — generated the training data backwards from clean Swedish text, gated it on a golden set of real dictations, and shipped it. The surprises were that the training data was the easy part, that my own generator contradicted itself, and that a full fine-tune can vanish inside a quantizer.

Lova is the AI planner inside Lovio: you tell it 'seat Anna next to Erik' and it does it, against your real guest list, budget and timeline. The interesting part isn't the model, it's how a turn stays tied to everything said before it without ever resending the conversation — retrieval instead of re-reading. Here is the whole machine.

For the past month I've been building Winston, a personal agentic OS that runs 24/7 on an always-on Mac: a conversational orchestrator I talk to over Telegram and a local web panel, backed by scheduled sub-agents that triage my email, work my task board, tend their own memory, and extend the system by writing new plugins. It uses no agent framework. The only agent dependency is the Claude Agent SDK, used as a primitive — and everything LangGraph or CrewAI would call 'the framework' is thirteen thousand lines of plain TypeScript around SQLite and the filesystem. This post is about why that shape, and what it buys.

On July 10 I got married, and every guest-facing system at the party — RSVP, seating, SnapShare, the slideshow, the toastmaster's presentation remote — was software I run in production for other couples every Saturday. We wrote about the wedding itself on lovio.se. This is the other half of that story: what it does to your engineering when the next release has a hard deadline, a hundred users in the room, and one of them is your wife.

My newest app went from an empty repo on May 19 to version 1.2.0 on the App Store in about nine weeks of nights and weekends. Not because I typed faster — because the agents did more than type. They built, ran the app in the simulator, tapped through the UI, read the logs, reviewed each other, and shipped the release. This is what Lykka is, what it runs on, and how that loop actually works.

I built Lykka's AI to say less, not more, and restraint turned out to take as much discipline as any feature does. Six and a half weeks, a hundred and forty-six commits, every one with an AI co-author in the trailer — the same restraint that shaped what the app says back to you shaped how I let it get built.

I shipped a habit tracker this year, built on one idea the whole category keeps getting wrong. The streak that resets to zero has quietly ended more of my own habits than the habits ever did. So I built the app that refuses to do that — and a forgiving model turned out to be much harder to build than a punishing one.

The unofficial App Store Connect CLI doesn't just happen to work with Claude Code. It was designed to be driven by an agent first and a human second. That ordering matters more than I'd assumed, and it makes me notice the few other CLIs that share the shape, and the many that don't.

A memory feature that prunes itself while you sleep is exactly what long-running agents need. The version that briefly shipped in Claude Code can also delete files without telling you, and silently stop running when it crashes. I want it back, with better manners.

The people shipping with agents in 2026 aren't the ones who can spell MCP. They're the ones who can read a five-year-old codebase in an afternoon and tell you which migration is keeping the team awake.

An honest log of one Saturday on Jobado: when I drove, when I steered, when I had three agent sessions running at once. Including the two times I had to undo the agent's work, and why I think that's still a win.

AI accelerates the typing. It does not accelerate the judgment. Most senior work isn't building. It's the careful undoing of last year's correct decisions.

Two versions of seniority. Both real. One is undervalued in 2026 hiring, and I want to make a small case for it.

Wedding-planning SaaS has a strange constraint: most failure modes are tolerable Monday morning. The one on the wedding day is not. Building for that asymmetry changed how I think about availability.

The most senior engineer in a room is usually the calmest. They've been wrong enough times to know the cost of dramatic certainty.

Most of the value I get from MCP isn't novel agent capability. It's the same connect-and-call ergonomics every codebase has wanted from its IDE for fifteen years.

A pragmatic, ordered checklist. Where the trust lives, where the load-bearing decisions hide, and the three files that almost always tell you the truth about a team.

An end-of-year list. Not what's new, but what I've quietly stopped saying. Including 'we'll fix it later,' 'just one more abstraction,' and 'they'll read the README.'