I’ve founded companies not because I had a clever exit strategy, but because I kept running into problems that nobody was solving honestly. In each case the same pattern emerged: take principled science, make it usable, and point it at something that matters. Here’s what that turned into.
Desupervised
Desupervised is where it all starts. It’s a Nordic AI technology company I founded in 2018 after leaving Blackwood Seven (which got acquired by Kantar). The core mission is deceptively simple: make advanced analytics accessible for everyday users. Not just data scientists, not just PhDs, but the people who actually need to make decisions every day.
We work across three areas. First, bespoke AI solutions where we build state-of-the-art models customized for specific business problems. Second, our platform Alviss AI (more on that below). Third, and this is the one closest to my heart, AI-driven drug discovery for rare diseases using Bayesian deep virtual screening. That last one is where my physics background really earns its keep.
As of today we’ve deployed over 2100 models across 15+ markets. Not bad for a company that started with a few physicists and a conviction that Bayesian methods were criminally underused.
Alviss AI
Alviss AI is our predictive analytics platform, built and operated under the Desupervised umbrella. It’s the productized version of everything I learned building marketing mix models at Blackwood Seven, except done properly this time.
Ok so what does “properly” mean here? Most companies model their business drivers independently. You have one model for marketing, another for pricing, another for churn, maybe another for brand. The problem is that these things aren’t independent. A price change affects churn, which affects lifetime value, which changes how you should think about acquisition spend. If your models can’t see those interactions, you’re optimizing in a fantasy world.
Alviss models ALL of your business drivers holistically. Everything that impacts your KPIs goes into one coherent model. And because we’re Bayesian about it (you saw that coming), every prediction comes with uncertainty quantification. You don’t just get “spend X on channel Y.” You get “spend X on channel Y, and here’s how confident we are about that recommendation, and here’s what happens if we’re wrong.”
We offer three tiers depending on how hands-on you want to be:
- Self-Service: you build and manage your own models on the platform.
- Hybrid: we help you set things up, you run them day-to-day.
- Full Service: we handle the data science, you focus on strategy.
The use cases range from marketing effectiveness and demand forecasting to churn prediction and pricing optimization. Basically anything where you need to understand what’s actually driving your numbers, not what your gut tells you is driving your numbers.
AI Alpha Lab
AI Alpha Lab is the other company, and it’s a fundamentally different beast. Where Alviss helps businesses make better operational decisions, AI Alpha Lab applies probabilistic machine learning to investment management.
The thesis is simple and I stole it from physics (as one does): investing is a deeply uncertain endeavor, yet most of the industry analyzes and forecasts with deterministic models. The result is, as we like to say, “the right answer to the wrong question.” We don’t ask whether something will happen. We ask what is the probability of something happening.
Our flagship product is a UCITS regulated, exchange-listed fund on Nasdaq Copenhagen: AI Alpha Lab Globale Aktier KL. The portfolio consists of 30 to 70 large-cap global equities selected entirely by our proprietary probability-based AI model. No human picks stocks. The model does. We launched the fund in November 2023.
The tracking error sits around 15-20%, which in finance speak means we’re genuinely different from the index. This isn’t a closet index fund that charges active fees for passive performance. It’s a concentrated, high-conviction portfolio that the model believes in probabilistically.
Beyond the fund, we also offer investment analysis subscriptions (including one through Nordnet Denmark), model portfolio delivery, and bespoke portfolio analysis where we run a client’s existing portfolio through our model to generate an uncorrelated alternative.
Now, the obligatory disclaimer because I actually mean it: past performance, whether actual or simulated, is not a guarantee of future returns. Investing involves risk. You can lose money. I say this not because lawyers tell me to (though they do), but because honestly acknowledging uncertainty is the whole point of what we do. If you can’t be honest about the risks of your own product, you shouldn’t be in the business of quantifying uncertainty for others.
The Common Thread
If you look at these three companies, the through-line is uncertainty quantification. In marketing, in drug discovery, in finance. The tools differ, the domains differ, the customers differ. But the philosophy is the same: if your model can’t tell you what it doesn’t know, it shouldn’t be trusted to make decisions.
I didn’t plan it this way. It just turns out that when a theoretical physicist starts building things, everything ends up looking like a posterior distribution. I’m ok with that.