Senior Data Scientist · Zürich
About me
Short bio
I started in physics: an MSc in Applied Physics at EPFL, then a PhD at ETH Zürich measuring how electrons leave matter on attosecond timescales. The lab taught me one habit I still rely on every day: a measurement is not a result until you have ruled out that you are measuring your own apparatus.
At eBay I brought that habit to business data over eight years and four roles, from data analyst in growth marketing to data scientist in eBay's Core AI group. Five of those years were in the global growth marketing organisation, working on experimentation, attribution, incrementality and personalisation. From 2024 I worked on eBay's conversational shopping assistant, taking it from capability evaluation to production.
Today I offer my data and AI expertise through my consultancy, Kasmi Data & AI. I work from Zürich, in English, French and German.
Experience and education
- 2024 to 2026eBay · Data Scientist II, Core AI, Agentic Science
The conversational shopping assistant: real-time customer context service, query autocomplete, LLM-as-judge evaluation and personalisation, shipped across web, app and messaging.
- 2022 to 2023eBay · Data Scientist II, Core AI, Marketing Science
User representation learning, the Recently Viewed Items homepage module, LLM-driven watchlist clustering, segmentation on the Customer 360 platform. Mentored two interns.
- 2020 to 2022eBay · Data Scientist, Marketing Science
Churn prediction and customer segmentation for the full customer base, owned end to end. Promoted within 18 months.
- 2018 to 2020eBay · Data Analyst, Global Growth Marketing Analytics
A/B experiments and incrementality studies for paid search and coupon campaigns across the US, UK and Germany.
- 2013 to 2017ETH Zürich · PhD in Physics
Ultrafast Laser Physics Group: attosecond photoemission and molecular relaxation dynamics.
- 2009 to 2013EPFL · BSc in Physics, MSc in Applied Physics
Selected results
What the work achieved
- 8×click-through on personalised starter prompts in eBay's shopping assistant, against a non-personalised control
- 89% · 92%answerability and resolution rate of the assistant, after evaluation cycles I designed with LLM-as-judge
- 98%of unsafe requests stopped by the LLM safety classifier I built as the assistant's first-pass gate
- 130Mcustomers scored every day by my churn model (F1 0.90, precision 0.89, recall 0.91) and segmentation pipeline
- 5% → 6%click-through among highly active customers on the Recently Viewed Items module I shipped to eBay's homepage
- 0.71 vs 0.18silhouette score of my graph-based user embeddings against raw features; first-author paper and pending US patent
Recognised with three eBay Spot Awards (2018, 2020, 2024).
For hiring teams
Roles I'm open to
Alongside my consultancy, I'm exploring my next long-term role: senior and principal positions where rigorous measurement or search and LLM systems sit close to the product.
- RolesSenior / Principal Applied Scientist, Senior Data Scientist, ML Scientist
- Strongest inExperimentation and causal measurement · personalisation and recommender systems · search and retrieval · LLM evaluation and agentic systems
- ToolsPython, SQL, PySpark · PyTorch, Hugging Face Transformers, scikit-learn, LightGBM, XGBoost · LangGraph, LangChain, RAG, PEFT/LoRA · Databricks, Airflow, MLflow, Docker, Kubernetes · GCP, AWS
- What I bringI've taken projects from the first question to production at eBay scale, with rigorous measurement, monitoring and quality checks along the way.
- LocationZürich, on site or hybrid; remote or hybrid for teams elsewhere
- AvailabilityImmediately
CV available on request.
Projects
Code
Public projects from my Udacity nanodegrees in data science and ML engineering.
- Disaster message classifier, from raw data to web appAn ETL pipeline into a database, a multi-label NLP classifier, and a Flask web app that sorts incoming messages into response categories for emergency teams.
- Credit card churn model as a tested Python packageIdentifies the bank customers most likely to churn, with logistic regression and random forest models, SHAP explanations, unit tests, logging and a command-line interface.
- Churn prediction for a music-streaming service with SparkFeature engineering on user session logs and a random forest churn classifier, built with PySpark.
Publications
Selected research
Industry research
- User Embedding and Community Detection by Contrastive Learning Augmented by a User-User GrapheBay internal AI conference First author
- US patent application US20250209484A1Pending First-named inventor
PhD, ETH Zürich
Attosecond band structure effect in photoemission dynamics from solid surfaces, ETH Zürich (2017). Best Student Talk Award, ATTO17 International Conference (2017).
Journal articles
Attosecond and ultrafast physics: 14 journal articles, cited more than 1,100 times. Full list on Google Scholar.
- Effective mass effect in attosecond electron transportOptica 4, 1492 (2017) First author
- Femtosecond single-electron pulses generated by two-photon photoemission close to the work functionNew Journal of Physics 17, 033008 (2015) First author
- Attosecond dynamical Franz-Keldysh effect in polycrystalline diamondScience 353, 916 (2016)
- Attosecond optical-field-enhanced carrier injection into the GaAs conduction bandNature Physics 14, 560 (2018)
Show 3 more publications
- Attosecond screening dynamics mediated by electron localization in transition metalsNature Physics 15, 1145 (2019)
- Light-matter interaction at surfaces in the spatiotemporal limit of macroscopic modelsPhysical Review Letters 115, 137401 (2015)
- Ultrafast relaxation dynamics of the ethylene cation C2H4+Journal of Physical Chemistry Letters 7, 1901 (2016)
Topics I'm exploring
What I'm thinking about at the moment
Physics intuition for learning dynamics
Whether ideas from ultrafast spectroscopy, such as relaxation timescales and symmetry breaking, can describe how neural networks learn and when new capabilities appear.
Evaluating AI output
What a trustworthy evaluation of LLM and agent systems actually requires, which parts are genuinely hard, and which are theatre.
Hybrid retrieval
Combining lexical and dense search with rerankers, and measuring honestly what each stage adds.