Ralph ToumaHome Work Machine learning research at Green-Communications
Role
Machine learning research at Green-Communications
As a machine learning research intern at Green-Communications in Paris (July to November 2025), I turned federated and clustered federated learning papers into working systems for weather forecasting and heterogeneous distributed networks, handling preprocessing, clustering, configuration generation, evaluation and performance analysis.
The brief
Green-Communications works on distributed network infrastructure, which makes it an unusually honest testbed for federated learning: the clients really are heterogeneous, and the convenient assumptions of a benchmark dataset do not hold. My work there was to take methods from the literature on machine learning, federated learning and clustered federated learning and find out what they do on real use cases, including weather forecasting and heterogeneous distributed systems.
From paper to running system
Most of the work lived between a published method and a result anyone can trust. In practice that meant data preprocessing, clustering, configuration generation, model evaluation and performance analysis, and then validating the outputs against the metrics they were supposed to hit, rather than against the metrics they happened to produce.
Comparison as method
Clustered federated learning is a claim about structure: that grouping similar clients beats treating them as one population. Testing that claim meant running the same problem through multiple experimental and training approaches and comparing them under identical conditions. I built the configurations, ran the comparisons and analysed where the differences actually come from.
Turning feedback into systems
Research feedback arrived as moving requirements. I converted it into scalable, testable solutions: improving experimental workflows so runs are repeatable, validating generated outputs so errors surface early, and adapting project components as performance targets and requirements changed.