I build the forecasting, optimization and causal-AI systems that turn frontier mathematics into board-level business decisions.
Research Collaborator, Stanford ICME & System X · Granted Patent Holder
I am a researcher and ML/AI executive — a PhD astrophysicist by training, now working across forecasting, mathematical optimization and causal AI. Alongside a research collaboration with Stanford ICME and System X, my work spans 30+ applied research projects across industrial, financial and regulatory domains.
Four fields, one translation problem: research into decisions.
Each area links to its projects — from open-source packages to working papers and early concepts, labelled honestly by status.
Peer-reviewable work, written for people who make decisions.
- 01
ForecastLens
What is a forecast actually worth — and in which regime?
>50%accuracy spread between market regimes, hidden inside a single averaged CRPS - 02
diffusim
How fast will a technology nobody has bought yet be adopted?
−25%forecast error (CRPS) versus nonlinear-least-squares baselines - 03
Supply Chain Digital Twin
Can you mathematically guarantee a supply chain won’t break?
1.29 vs 2.26test MAE of the GNN risk model against an additive baseline - 04
Cat Pathfinder Optimizer
Plan like a cat: avoid what matters, remember what failed.
−38%replans with the full model, at an equal 100% success rate
How the work gets done.
Five habits that show up across every project in the portfolio — each one traceable to real code, not a slide.
- STEP 01
Start from the decision
A forecast is only as good as the choice it improves. Work begins with the decision it has to support — buy now or wait, reroute or hold, allocate or not.
Seen in: ForecastLens
- STEP 02
Validate against known truth
Every method is checked on data where the answer is known before it touches real data: synthetic presets, independent numerical schemes, explicit causal ground truth.
Seen in: diffusim
- STEP 03
Prefer structure you can audit
Causal graphs with provenance on every edge, domain-neutral contracts and interpretable operators — so a result can be traced, questioned and trusted.
Seen in: chronocause
- STEP 04
Break it on purpose
Leak-safe walk-forward backtests, operator ablations, regime-by-regime scoring — and negative results documented instead of hidden.
Seen in: Supply Chain Digital Twin
- STEP 05
Report what it is worth
Results in the units decision-makers use: savings per unit, replans avoided, interval coverage that actually matches its promise.
Seen in: Cat Pathfinder Optimizer
The same models that pass a peer review also run the demand plan for a multi-billion-euro materials group.
Not where I worked — what it changed.
First team member of a materials-industry corporate venture function. Built a demand-forecasting platform now protected by a granted patent (LU103146), plus a CO2/sustainability platform and the group’s first digital-twin and replenishment initiatives.
Built a RAG and causal-intelligence engine that turns market news into queryable, causal signals. Optimization models now identify each customer’s optimal purchase timing and hedge ratio, scaled across thousands of features per commodity.
Built forecasting and mathematical optimization models to strengthen production planning and supply-chain decision-making.
Supported stress-testing of global loan and mortgage portfolios within liquidity and crisis-planning frameworks aligned to EBA and Basel requirements.
Published research, industry writing and keynotes.
- 2026Working Paper
From Hourly Noise to Daily Signal: A Regime-Aware, Statistically Validated Ensemble for Brent Crude Oil Forecasting
Preprint in preparation
- 2025Article
Wie KI Rohstoffmärkte berechenbar macht
Beschaffung Aktuell
- 2023Keynote
Decoding Tomorrow: The Evolution of AI in Demand Forecasting
KI Park × Google × Pacemaker, Google Office Berlin