An engineer who reads the papers and then ships the thing

The short version

I build machine learning systems that hold up outside the notebook. My research work is in federated and clustered federated learning: training models across clients that hold their own data and never hand it over. My engineering work is in the less glamorous half: preprocessing, clustering, evaluation, configuration, and the dashboards and storefronts that put a result in front of someone who has to act on it.

How I got here

I started on the web side. Since May 2023 I have been a developer at BYND Network, building Shopify storefronts and custom apps for more than 15 clients, along with the performance dashboards their teams use. That work taught me something a course cannot: that an insight only counts when a non-technical stakeholder can act on it. Applying data analysis to storefront behaviour improved conversion by 30%, and that number did more to shape how I think about modelling than any benchmark.

A backend bootcamp and internship at Eurisko in 2024 filled in the systems half: TypeScript, Node.js, Express, MongoDB, REST API design, modularity and error handling. Before that, freelance work for Appen had me evaluating how well virtual assistants answered real queries and structuring business data for AI training sets, which is where I first saw how much of machine learning is actually data quality.

From July to November 2025 I was a machine learning research intern at Green-Communications in Paris, working on federated and clustered federated learning for weather forecasting and heterogeneous distributed systems, taking methods out of papers and finding out what they do when the clients are genuinely different from each other.

What I am interested in

Learning under constraint. Federated learning is interesting precisely because the easy option (pool everything centrally) is off the table, and what is left is a real question about heterogeneity, personalization and honest measurement. The same instinct shows up in my retrieval work: a RAG assistant that cannot cite its source is a confident guess, and I would rather build the version that has to show its evidence.

How I work

Compare before you conclude. Every project on this site has a baseline attached to it, because a model with no comparison is a claim with no evidence. I validate outputs against the metrics they were meant to hit, track error where it can hide (per client, not just globally) and write the result down in a form someone else can check.