Product Manager · Data & AI · 0→1 builder
Product manager with 12 years of experience and hands-on data practice: I define KPIs, build predictive models myself, and turn raw data into decisions. This page is where I keep the things I build. Hobby and portfolio projects, the kind that don't fit neatly on a CV.
Full career history and experience on LinkedIn →
Deliberately spread across applied AI, data science, and interactive storytelling. The through-line isn't a single stack: it's taking each one from a blank page to something real, finished, and genuinely pleasant to use.
Job search for Switzerland that reads every posting so you don't have to. It scores each one on skill fit against your CV and interest fit against a page you write about what you want next, then drafts a tailored CV and cover letter in the posting's language. Runs daily in the cloud; released as an open-source template.
A personality-driven AI agent that embodies a beloved plush bear, with persistent memory across sessions, LLM tool-use for durable facts, and both CLI and Streamlit interfaces.
A retrieval engine over Marcus Aurelius' Meditations: you describe what's troubling you, it returns the fitting passages with citations, in Marcus's own words rather than an LLM's. Built as a testbed for advanced RAG: routing, query transformation, reranking, and an eval harness strict enough to say which of them actually earned its cost.
End-to-end referral-program analysis on synthetic data: a referral-cascade graph, a two-proportion z-test, and an interpretable XGBoost model (Weight-of-Evidence + SHAP). A privacy-safe adaptation of real production work.
Capstone of Google's Advanced Data Analytics certificate: predictive modeling of employee attrition, from exploratory analysis through to a deployed-quality model.
An illustrated guide to dinosaurs and prehistoric life, localized in six languages, for young explorers and the young at heart.
Short, practical write-ups of what I learned building things, with the ideas general enough to carry to the next project. Grouped as collections; details generalized.
A five-part series on taking a chain of LLM agents from "a different answer every run" to reproducible: contracts between stages, single-sourced facts, deterministic output, graded test data, and a drift eval to prove it holds.
A two-part series on what it takes to get from certificate exercises to a first model for a real use case: class imbalance, Weight-of-Evidence encoding and the target leakage a pipeline prevents, then judging the score, knowing when to stop tuning, and reading a SHAP summary plot.
All writing → Field notes index