The unorthodox journey of an AI engineer

This week I asked Claude what it could see of what I've worked on this year. I thought I had started using it in February. It found my first session: 15 January 2026. It also found the first thing I typed after it answered:

"Just wanted to see what you were able to see."

Nine months apart, I asked the same question. Everything in between is below. It isn't the path you usually read about, from computer science degree to job to startup. It's the path of someone who spent fifteen years owning a number and kept building the tools to see it.

Before AI: building where the number lives

My career was on trading desks in ad tech, mostly in London. The job was always a number: a campaign's margin, a desk's P&L, a client's spend. I kept building the tools that made that number visible, usually without anyone asking me to.

A P&L system I built at one employer was in daily use for eight years, long after the team that asked for it had left. Reporting I automated for about 600 accounts was later taken up by two other countries. None of it used AI.

That matters more than anything that follows. My tools changed every few months. The instinct, to find the number and dig out the data behind it, never did.

2023: a board of directors made of chatbots

My first AI build, in August 2023, was a board of directors. It was a Google Apps Script tool that called OpenAI several times, gave each call a different director's role and let them argue about a decision before returning their advice.

It was a toy. It also gave me the idea behind almost everything I've built since: one model, given different roles, can be made to disagree with itself.

Then came LifeLeap, an AI life coach. I had worked with a human accountability coach, so I taught the AI, with OpenAI underneath, to talk like her.

I built LifeLeap in FlutterFlow, a visual app builder. Ironically, that's where I learned code logic. Every screen, action and data call was drawn out and connected, so I could see how an app actually fits together before I wrote much code myself. LifeLeap also took me through the parts nobody makes videos about: getting an app into App Store Connect and Google Play, and setting up subscriptions.

I spent about £25,000 on design before I shut it down. It's the most expensive lesson I've had: a polished app isn't the same as a product. It also taught me to kill things, which I have done many times since.

2024–2025: copy, paste, ask again

For a long time, AI meant ChatGPT in one window and my code in the other. I pasted a function in, asked for a change, pasted it back and saw what broke. Tiny changes, one at a time.

In August 2025 I moved to Cursor, where the AI sits inside the editor and can see the whole project. It was the first time it felt less like asking and more like pairing. Between August 2025 and January 2026 I built these in Cursor:

• crypto analytics tools
• Telegram bots
• a tool that turns video files into ad tags
• a Chrome extension that strips the feed out of social media

Lovable: where things actually shipped

Alongside Cursor I used Lovable, where you describe an app and it builds it. That's where my work started reaching other people:

• AudienceScan, my web3 analytics product, got its front end there: more than 2,100 changes since July 2025.
• A health log that our family uses every day.
• A bedtime-story app for my kids, voiced through ElevenLabs. When a story ends, an AI pass picks out the characters and events, and the next story remembers them.
• The first version of Spot Fitness, an app for Stockholm's outdoor gyms.

Lovable taught me that shipping something small to real people beats planning something big.

January 2026: from writing code to directing it

Claude Code was a different kind of step. It doesn't suggest the change, it makes it: it reads the project, runs the commands and deploys. My first session, on 15 January, was a small dashboard on top of an ad platform's API. Within a week I was rebuilding AudienceScan's whole backend on Google Cloud.

In January I also installed a tool that keeps Claude working on its own, in a loop. On 22 January the usage counter shows 26,000 messages in two sessions. I didn't type them.

On 27 February I asked Claude a half-serious question: what if I hired you full time and checked in every hour? I didn't do anything with it then. Hold that thought.

Spring: building things nobody asked for

In March I built AgentSpawner, a browser extension where you describe a task and an AI builds the workflow. In April it was a LEGO-based wargame, just for fun.

From the end of April to June I built a trading platform. Nine accounts traded in parallel on an exchange, and a BigQuery viewer mined the trade history for rules. It made more than 50,000 live trades at minimal stakes, and it broke even. The edge was real on paper, but fees and slippage ate it. I shut it down. What stayed was the discipline: test against real outcomes, not backtests, and believe the result.

Something smaller from April turned out to matter more. I decided to offer to build things for friends' companies for free, in exchange for case studies. The next day I built a first test for a Swedish company. In October the same company booked eight paid sessions.

Summer: an update a day, and a patio

In June I moved Spot Fitness from Lovable to Xcode as a native iOS app. In July I shipped almost an update a day, while also using AI to plan a patio at my parents' summer house. By the end of July I had App Store screenshots. Version 1.0 was finished in August, with Apple Health and cues on the watch.

I'm still working on it. Claude Code can now drive the iOS simulator, which I found out while building an app for a law firm, so it tests the app itself.

The job hunt that built a company

At the end of July I started looking for a job. I'm saying that plainly, because it shaped everything that came after.

I wrote a build log of every project since 2023, which is where half of this post comes from. One job ad asked for dbt, a tool for structuring and testing data. I didn't know it, so I asked Claude to build me a course. A few weeks later dbt was running in production in my own retail data platform.

In September I applied to an AI company that works with forward-deployed engineers: people who sit with the customer, build inside the customer's systems and hand the result over. A recruiter explained how they charge: a platform licence, a yearly fee for what they've built and keep running. I had simply never considered that structure of work, and it opened up a world I hadn't had access to. The same day it became Promtech's model. What we build for a client is theirs; what runs every day is ours.

Another process, with a large consultancy, is still going. Getting this far has already told me something I needed to hear. A large company has a job called AI engineer, and the people hiring for it think I can do it.

A job hunt is market research you can't get any other way. Every interview showed me what someone would pay for.

September: the first invoice

On 25 September I sent my first invoice for a platform I had pitched and built as a forward-deployed engineer. It was a system for a crypto company that connects their LinkedIn ad data to the companies behind the clicks, with a managed-service fee on top of their ad spend. I wrote the message telling Claude in capitals.

October filled up:

• a pilot with a fashion brand on how their products sell at retailers
• finance automation for a Swedish company
• a case and calendar app for a law firm, on web and iOS
• a pipeline that turns my client calls into notes and follow-ups

On 22 October I built the thing from February: an app on my Mac where Claude acts as my manager. It checks in through the day and steers me towards what makes money. That was eight months after the half-serious question.

What I'd tell someone starting now

• The tools will change. Your instinct won't. OpenAI, FlutterFlow, ChatGPT, Cursor, Lovable, ElevenLabs, Kling, Claude. What carried over was knowing which number mattered.
• Start visual if that's what makes it click. FlutterFlow taught me how an app fits together before I wrote much code.
• Build where the number lives. Margin, spend, lost sales, pipeline. Not generic AI features.
• Kill things. LifeLeap, the trading platform, offers I couldn't deliver. Each one made the next thing better.
• Ideas come back. The board of directors became agents. The hourly check-in became my manager.
• Treat a job hunt as research. Mine showed me the business model I now run.

The timeline

Back to the question

I wrote this post the way I build most things now: I asked Claude what it could see, then filled in what it couldn't. It could see the sessions and the code. It couldn't see the board of directors, the patio or the recruiter call.

That's still the split. The tools see more every month. Knowing which parts matter is the job.