Consulting
I take on a small number of projects at a time, each scoped around one decision or one system. You work with me directly, from framing the question to handing over something your team owns.
“Is our marketing spend actually driving growth?”
For growth, marketing and finance teams
Design and run incrementality tests and experimentation programmes, build attribution you can reason about, and help your team read results honestly, so budget moves to what causes sales rather than what correlates with them.
“It works in the demo. Will it work for every user?”
For product and engineering teams
Evaluation harnesses that gate releases, retrieval and ranking improvements measured offline and online, and tracing that ties every answer to the prompt, model and context that produced it.
“We have years of process data and a lot of tacit know-how.”
For R&D and engineering leads
Models and optimisation grounded in the physics of the process, for laser, photonics and other physical systems, from someone who has aligned the lasers herself.
How engagements work
Two weeks, fixed fee. I look at your data, your setup and the decision you need to make, and hand back written findings with a prioritised plan.
Six to eight weeks on one outcome: a measurement programme running, an evaluation harness in production, a model your team owns.
Optional monthly support with capped days and a written response window, for teams who want a senior partner on call.
What you can expect
Every recommendation comes with how we would find out it is wrong.
Assumptions, error bars and limits are written down, not buried.
Code, documentation and a proper handover, in your stack.
You work with me, not a junior team, in English, French or German.
Selected work
eBay · Growth marketing
Five years in eBay's global growth marketing organisation, first in analytics, then in data science: A/B testing, channel attribution and incrementality measurement for paid search and coupons. The question behind all of it was which spend actually caused the sale.
eBay · Retention
A LightGBM churn model (F1 0.90), exported to ONNX and run as a daily batch over the full user base, alongside a customer segmentation that labels the same users every day for targeting.
eBay · LLM systems
LLM intent extraction feeding structured product search, with an evaluation suite that gates every release, versioned prompts and models, and traces joined to the platform's own tracking.
eBay · Personalisation
An autoencoder over a co-purchase graph combined with a Siamese network trained on triplet loss, run deliberately as a feasibility study to decide whether graph signals were worth productionising.
ETH Zürich · Research
PhD measuring photoemission time delays of electrons on the attoline beamline, plus gas-phase work on the ultrafast relaxation of the ethylene cation. Co-author, J. Phys. Chem. Lett. 7, 1901 (2016).