Hannes Zetterblom

2014-2017

Sales, chosen as training: cold and warm calling for a security firm, ad sales at Hitta.se, then sales manager with hiring responsibility at a meeting-booking firm.

2016

First company at 21, a bracelet brand sold online. It broke even and taught me the basics of Shopify, supplier negotiation, SEO and paid social; after a year I closed it for the betting business.

2016-2025

My own quantitative betting business, my main work from 2017: Sweden, Norway and Canada, at times in parallel, with a team of up to seven and a monthly P&L. Sports until 2021: volume, hedged and arbitrage betting. The best year, 2018, netted more than 4.5 MSEK; Sweden's new gambling act of 1 January 2019 limited bonuses to one per player and operator and closed most of that edge. I rebuilt the business around slots, from 2021 to 2024: Excel models on data I collected at scale mapped the variance of nearly every provider's games, and Monte Carlo simulations priced each bonus against RTP and chose between high and low volatility. Both from 2024. It ran alongside KTH (2018-2019) and Clar and gave me complete freedom: I lived abroad for several years, traveled widely and pursued my own interests.

2021-2023

Clar, a fintech loan broker on four continents.I set up its Italian and Portuguese companies, from choosing the law firms to recruiting the country launchers, launched clar.co and wrote its launch copy, ran market-entry analyses for six countries, screened acquisition targets, and sat in on financing rounds and IPO preparation.

2025-

Engineering Mathematics at KTH, for the capability rather than the credential, alongside the systems below.

In April 2023 I pitched a gambling-affiliate business to a potential backer. One slide suggested that ChatGPT-class models had made much of an established competitor's 52-person product team less useful. The pitch was not funded. In 2026 I built the business anyway, with AI agents where that team would have been.

The result is oddslogik.se, a Swedish review publication for the licensed gambling market, where one wrong claim is a regulatory problem. It has over 80 pages, a decision log past 360 entries, and a doctrine file that every agent reads before it touches anything.

For a friend's services company I built a 4 230-row prospecting CRM with agents, then set a separate model instance to audit it over four sessions. It found about forty defects. A dozen were the builder's own rulings, accepted as correct: the kind of error a system cannot see in its own output. Nothing I build ships without that audit now.

TimeBox Stats is a Garmin watch app I built in Monkey C. A session started on the wrist syncs through the phone to Postgres, a public dashboard and the record below. Code, architecture guide.

This site runs the same way: its decision log is public, and the agents and I both write to it.

Every focused session I log on the watch reaches this page overnight. This is every day since 7 April 2026.

612 hours across 1 131 sessions, logged on 152 of 213 days; 15.7 hours in the last seven days. Days without a session are shown, not skipped.

Through 5 Nov; updated 6 Nov 10:51 CET.

• The first question is always: where is the leverage?

  Price against value, perception against reality, effort against payoff. The gap is the work.

• Read the rules more carefully than the people who wrote them.

  Most systems pay the person who does.

• Understand the generator, not the answer.

  Learn the mechanism that produces the answers, and most of them come free.

• Act on calibrated uncertainty.

  Waiting for certainty the domain cannot supply is a way of never deciding.

• Research that does not change a decision is procrastination with a good reputation.

• Discipline is a state, not a trait.

  Mine appears when progress is visible and the stakes are real, so I build environments that make both true. The heatmap is one.

• A high floor beats a high ceiling.

  Intensity is not consistency. The week you can repeat is the one that compounds.

• Courage is not the scarce input. Downside architecture is.

  Maximize compounded return subject to survival. Every rule I keep was paid for, so the set stays small enough to obey.

• A green build is not a deploy.

  Verification is a separate step, done by an instance that did not build. An instrument that hides its gaps is decoration.

• Decisions close with three lines: decision, rationale, next action.
• Every material choice gets a numbered log entry before the work starts.
• Plan first, halt, then write. Layout changes are shown before they ship.
• An agent's summary is a claim; the diff and the build are the evidence.
• Up to four sessions run in parallel, partitioned by file ownership, so nothing collides.
• "Redo it, do it right" is a normal sentence, in both directions.

I want 20 focused hours a week inside a small AI-native team. To start: one scoped build of 8-12 weeks, on something you already know you need, at a day rate. If it works, we continue at 20 hours a week.

A first build could be an enrichment pipeline, an agent workflow with a verification gate, or a data model for work your team still does by hand.

Scope a build