Bayesian neural networks
Deep learning with calibrated uncertainty. Predictions come with error bars, and the model tells you when it is out of its depth.
Bayesian machine learning · Sweden
AI that knows what it doesn't know.
We build machine learning systems that quantify their own uncertainty, from Bayesian neural networks to bespoke probabilistic models. That way you can make decisions you can defend.
Click to add data: model uncertainty shrinks. Stack conflicting points at one spot: noise grows.
What we do
Probabilistic modelling
Hierarchical models, Gaussian processes, state-space and time-series models. We pick the structure your problem needs, not the biggest one available.
Forecasting & decisions
Full predictive distributions feed straight into decisions such as pricing, planning, allocation and risk, so you optimise expected outcomes rather than point guesses.
Model audits
Already have models in production? We check calibration, robustness and failure modes, and tell you where the confidence is unearned.
Selected work
Media · Distribution
Delivering the magazines that actually sell
A Nordic media group
- Challenge
- The client delivers magazines to kiosks across its network. Every unsold copy comes back as a return, which is costly to handle and wasteful to print in the first place.
- Approach
- Bayesian prediction models for every kiosk, forecasting how many copies will actually sell, so delivery quantities are set from data rather than guesswork.
- Outcome
- 15–20% fewer returns: lower costs, and less paper printed only to be thrown away.
Real estate · Valuation
A price estimate for every home in Denmark
A Nordic real estate chain
- Challenge
- The client needed one consistent way to estimate what a home will sell for, anywhere in the country.
- Approach
- A Bayesian house pricing model that gives every estimate a credible range, not just a single number.
- Outcome
- A nationwide valuation engine covering all of Denmark, now offered by the client to its own customers.
Engineering · Robust design
Acting CTO for a robust design company
RD8 · founder's engagement
- Challenge
- RD8 helps engineers create designs that tolerate real-world variation, so products last longer and don't break. That makes them better for customers and for the environment.
- Role
- Michael Green served as RD8's acting CTO, leading its technology.
- Outcome
- RD8 launched its robust design platform, bringing longer-lasting designs to its customers.
How we work
-
01
Start from the decision
What will change depending on the answer? That defines what the model must get right, and what uncertainty matters.
-
02
Encode what you know
Domain knowledge goes in as priors and structure. The model learns more from less data, and its assumptions are explicit and testable.
-
03
Ship, monitor, update
Calibrated models in production, checked against reality. When the world shifts, the posterior moves with it.
Green is in the name
Part of it is the founder's surname. The rest is a conviction that good machine learning shouldn't burn more than it needs to.
- Right-sized models. A well-specified small model often beats a giant one and costs a fraction to train and run.
- Data efficiency. Priors and structure mean fewer experiments, less data collection and fewer retraining cycles.
- Honest about compute. We weigh energy cost alongside accuracy when choosing between approaches.
- Problems that matter. We're especially keen on work in energy, climate and sustainability.
About
Green AI is a Swedish company founded by Michael Green, PhD. We build Bayesian machine learning for clients across the Nordics. Michael writes about Bayesian modelling and AI at drmike.xyz.
Have a problem with uncertainty in it?
Most interesting problems do. Tell us about yours.
hello@nextai.green