Lycoriolis
I build models, then argue with them until they hold.
I work on machine learning, mostly on the unglamorous half: how a model is framed, measured, and made trustworthy enough to use.
Good results are easy to produce once and hard to carry. Most of my time goes into the difference between those two things.
Off the clock I am usually reading something old, or taking apart a game engine to see how it decides.
“The limits of my language mean the limits of my world.”
Ludwig Wittgenstein
Make the system legible.
A model earns its keep when the person using it can see the assumptions, the failure modes, and the next move.
Be suspicious, politely.
Most of the work is not the model. It is noticing that a metric is flattering itself, or that a clean number came from a dirty join.
Leave a trail.
Notebooks, pipelines, and writing good enough that a stranger can disagree with the work on its merits.
Domain-Adapted Language Models
Fine-tuned a Qwen language model on aviation technical problems and built a GPT-2-scale transformer from scratch to study tokenization and training dynamics.
Monte Carlo Tree Search Engine
Implemented an Othello engine powered by Monte Carlo Tree Search (MCTS), exploring sequential decision-making and search heuristics under time constraints.
Vehicle telemetry & predictive maintenance
Built high-volume vehicle telemetry pipelines for predictive maintenance, anomaly detection, and fleet-level operational insight, from sensor ingestion to model-ready features.
Vehicles are unforgiving data systems: late events, missing sensors, and a decision that has to be safe before it is clever.
Healthcare Data Engineering
Designed and deployed a medical data pipeline for a hospital service, standardizing ingestion, validation, and downstream clinical reporting.
Working with clinical data sharpened my approach to data governance and failure modes, alongside plenty of late-night pipeline debugging before deployment.
Competitive ML & Hackathons
Two-time Hi!Paris hackathon participant and active Kaggle competitor, building rapid tabular and deep-learning solutions under tight compute and time budgets.
Machine learning, end to end
Framing, statistical and causal inference, evaluation, and the engineering that lets a model live outside a notebook.
- Problem framing and evaluation under shift
- Statistical and causal inference
- Reproducible pipelines and deployment
- Clear interfaces for operators
Automotive Big Data & ML
High-volume vehicle telemetry and sensor streams for predictive maintenance, driving safety, and fleet-scale operational insight.
- Telemetry & sensor ingestion pipelines
- Predictive maintenance & anomaly detection
- Driver behavior & fleet analytics
Clinical Data & Decision Support
Dependable data engineering and modeling for hospital environments, designed around strict privacy, auditability, and clinician oversight.
- Clinical data transformation pipelines
- Explainable diagnostic & service analytics
- Privacy-aware, accountable deployment
Different stakes, same discipline: know what the data can support, and say so plainly.
Statistical & Causal ML
Models grounded in solid evaluation metrics, causal reasoning, and clear assumptions rather than benchmark chasing.
Human-Centered Decisions
Interpretable systems designed to sharpen human judgment and keep accountability with domain experts.
End-to-End Infrastructure
Reproducible data pipelines and cloud deployments that stay maintainable after the prototype stage.
Clear Technical Interfaces
Fast web tools and technical writing that make complex model behavior legible to engineers and operators.
If the signal is messy and the decision has consequences, I want to hear about it.
Pushing transformer architectures from raw BPE merges to domain fine-tuning makes their strengths, and their brittle spots, much clearer than treating them as black boxes.