I build tools when the existing ones don’t do what I need, or when they do it without respecting uncertainty. Here’s a selection of software I’ve created or helped build.
Proprietary
Alviss AI
Alviss AI is the marketing science platform I built together with the team at Desupervised. It’s the productized version of everything I learned building Hamilton AI, except done properly: all your business drivers modeled holistically, every prediction with uncertainty quantification, and an architecture that scales from self-service to full-service engagements. It covers marketing effectiveness, demand forecasting, churn prediction, and pricing optimization across 15+ markets with over 2100 models deployed.
AIChemy
AIChemy is an AI-driven early stage small molecule drug discovery platform I built at Desupervised. It uses equivariant probabilistic graph neural networks to predict molecular properties, making it possible to screen compounds faster and with honest uncertainty estimates. This is where my physics background earns its keep the most: the equivariance ensures the model respects the symmetries of molecular geometry, and the probabilistic layer tells you when it’s guessing. Built for rare diseases that the market won’t fund.
AI Alpha Lab Decision Engine
The proprietary decision engine behind AI Alpha Lab’s UCITS regulated fund on Nasdaq Copenhagen. It’s a Bayesian neural network engine that integrates financial market knowledge based on decades of data to select a concentrated portfolio of 30 to 70 global equities. Every prediction comes with calibrated uncertainty, which is the whole point: if you can’t quantify what you don’t know, you shouldn’t be managing other people’s money.
Hamilton AI
Now part of Kantar Lift ROI
Hamilton AI was the marketing mix modeling platform I built as Chief AI Officer and cofounder at Blackwood Seven. It was one of the first systems to apply Bayesian methods at scale for media effectiveness measurement, modeling the full marketing ecosystem holistically rather than treating each channel in isolation. When Kantar acquired Blackwood Seven, Hamilton AI was absorbed into their product suite and renamed Lift ROI, where it continues to power marketing optimization for global brands.
Open Source
Borch
Borch is a deep universal probabilistic programming language I built together with Johan and Lewis in the early days of Desupervised. It’s the foundation that Alviss AI, AIChemy, and the AI Alpha Lab Decision Engine are all built on. The idea was to make Bayesian neural networks as easy to work with as regular neural networks: you define your model, Borch handles the uncertainty. Over 3000 commits and 12 releases later, it’s still the engine under the hood of everything we do. Licensed under Apache 2.0.
NeuralNetHack
GitHub · C++
A fast, lightweight C++23 library for training and evaluating ensembles of multi-layer perceptrons with zero external dependencies beyond an optional BLAS library. I started this back in 2004 and it’s been my workhorse for understanding neural networks from the ground up.
It supports multiple activation functions (Sigmoid, TanH, ReLU, Leaky ReLU, ELU), optimizers (SGD with momentum, Adam/AdamW, L-BFGS), batch and layer normalization, dropout, and ensemble learning with bootstrap and cross-validation. Everything you need to train and evaluate neural networks without dragging in half the internet as dependencies.
dammmdatagen
GitHub · R
An R package for generating realistic synthetic marketing mix modeling (MMM) datasets where the ground truth is known. If you’ve ever tried to validate a marketing mix model, you know the problem: you never know the true effect of each channel. This package solves that by letting you generate data with known relationships.
It produces covariate data (weather, economic indicators, competitor spending), response variables, online and offline media metrics, and macroeconomic indicators. Useful for benchmarking MMM approaches, teaching, and sanity checking your modeling pipeline before pointing it at real data.
EvidentialFlux.jl
GitHub · Julia
A Julia package built on Flux that implements Evidential Deep Learning for estimating both aleatoric and epistemic uncertainty in a single forward pass. This is the kind of thing I keep coming back to: if your model can’t tell you what it doesn’t know, it shouldn’t be making decisions.
It supports deep evidential regression using Normal-Inverse-Gamma distributions and deep evidential classification with Dirichlet distributions. You get uncertainty quantification without the computational overhead of Monte Carlo dropout or deep ensembles. One forward pass, two types of uncertainty. That’s the deal.
BASE
base.thep.lu.se · Java
BASE (BioArray Software Environment) is a free, web-based database and LIMS for managing sequencing and microarray experiments from sample to analysis. I was a developer on this project during my time at Lund University. It grew into a widely used platform in genomics research, handling everything from multi-user access with role-based permissions to a plugin architecture that let labs extend it for their own workflows. Licensed under GPLv3 and still actively maintained.