Lab

Small experiments, running in your browser.

Things that do not deserve a full case study, and things that are easier to show than to describe. Nothing here sends data anywhere — it all runs client-side.

Interactive

Explore an industrial dataset

5 000 synthetic assembly records, generated in your browser from a process model I wrote — no real company data, no server, no tracking. Pick a feature to see how it separates good parts from scrap, then move the sliders and watch a calibrated classifier decide.

Feature distribution

OKNOK
n
mean
std
NOK rate

Live prediction

gradient boosting · calibrated
p(OK)


How this one works

What you are actually looking at

The data is synthetic — on purpose

Everything I build at work sits on confidential process data, so none of it can go on a personal website. Instead, this dataset is generated in your browser from a process model I wrote: a nominal bore diameter with realistic scatter, press force variation, tool wear accumulating over a production run, and ambient temperature as a weak effect.

A seeded pseudo-random generator means every visitor sees exactly the same 5 000 records, which makes the numbers on the page reproducible rather than decorative.

The decision rule is the interesting part

The classifier outputs a calibrated probability, and the routing rule has three outcomes rather than two: assemble, re-measure, or reject. The thresholds come from the relative cost of each error — which is why they are not 0.5.

This is the part that most portfolio demos skip, and it is the part that decides whether a model is useful on a line. Move the sliders into the tails and watch a part change category.


Also in progress

Currently poking at

Side experiments that will land here as they become presentable. This page is designed to grow — a portfolio that is finished is a portfolio that goes stale.

  • Anomaly detection benchmarks

    Comparing detector families on synthetic process signals with metrics that reward early, single detections instead of point-wise overlap.

  • Local LLM evaluation harness

    A small framework for scoring open-weight models on domain-specific prompts — the public leaderboards do not answer the question I have.

  • Surrogate models for FEM

    Fitting fast surrogates to simulation output so a design space can be explored interactively rather than overnight.

  • Sports analytics

    Training load, pacing and ski-touring route data. Mostly an excuse to apply time-series methods to data I care about personally.


Contact

Currently looking for the next problem worth solving.

Industrial AI, manufacturing data, computer vision, generative AI — or anything where engineering and machine learning have to meet in production. If that sounds like your team, I would like to hear about it.

Mauren, Liechtenstein