I'm a machine learning engineer. I build things that learn, then measure whether they actually did, which is where most of the fun (and most of the bugs) live.
By day I build forecasting and AI systems, lately in energy, finance and luxury retail. By night I go back to old projects to check whether past me was right. Sometimes he was.
Four years of forecasting, data pipelines and AI, mostly where energy meets money.
A B-spline regression for how the weather moves energy volumes, in SQL and Python. It beat the old model by 3% in A/B tests and led to about £600K a year in savings. Also: Power BI dashboards and data guidelines for the trading and hedging teams.
For one of Romania's largest banks: agents read mortgage dossiers (scanned contracts and land-registry extracts, stamps and handwriting included) and draft the legal opinion a jurist would write by hand. I owned the ingestion (layout parsing with a vision-model fallback, over 95% extraction accuracy) and the evaluation framework, where LLM judges score each opinion against the jurists' own.
Visual similarity search on Azure AI across 500K+ products, demand and order forecasting per store from each new article's visual twins, and store-capacity benchmarks with bootstrap intervals and trend tests, shipped as FastAPI microservices with Docker on Databricks and PySpark.
Time-series anomaly detection on high-frequency battery data for an electric aerospace startup, pitting LOF against LSTMs and autoencoders, and a co-written Python package for 8 battery cyclers (cleaning, slicing, feature extraction, plots) that cut preprocessing time by over 83%.
Gaussian-process models trained on HPC to estimate battery state of health, blending lab data with PyBaMM simulations so fewer experiments are needed.
An 8.8-million-title IMDb dataset on BigQuery and a first BERT model for stereotypes, which I later took apart and rebuilt properly (the story).
Old repos, rebuilt so that they run, measure themselves and say honestly which parts are mine.
How film and TV cast, write about and label women and social groups: 13 studies on IMDb's catalogue, TMDb and 3 million plot summaries, 1920–2026.
Robots that learn a maze on their first run: my 2022 coursework explorer, a D* Lite map planner and a route prover that guarantees the shortest second run, benchmarked on thousands of seeded mazes.
A procedural terminal adventure with a built-in autopilot trained with CEM and PPO. Grab the controls any time; dodge the patrols, collect the crates.
Games, fractals and image processing in C, including a Tetris AI with cross-entropy-trained weights and reproducible benchmarks.
Tools for cleaning and slicing battery-cycler data. Being refreshed right now.
Communities, cars and the occasional conference I've led or helped build.
Ran what was then the UK's largest student-led Data & AI conference (100+ attendees; my talk was on AI history and NLP in films), tripled sponsorship to £5K from companies such as Google, Jane Street and WorldQuant, and kept 12 sub-teams and 2,000+ members going.
Started two chapters; at Warwick, 13 events (ML, Hash Code, Flutter, design sprints, careers) and 520 attendees.
A telemetry analytics pipeline in R and MATLAB whose insights helped cut lap times by over 15%, and a six-person team prototyping an autonomous control system for a Formula Student electric car.
Part of the team behind the Social Impact Award România 2020 winners.
Mostly sport, food and looking for good walls.
I've played my whole life, very seriously at Warwick, and this year I joined a team in an amateur league in Bucharest. Two wins so far, and I'm claiming a fair share of the credit.
I love it, but I'm short of a regular partner and the time to practise. Bucharest players, get in touch.
Current goal: deadlift two of my cousins at once. They're both heavier than me.
I used to cook and bake a lot (and post the results on Instagram; I might start again). These days it's coffee at home with very basic latte art, surrounded by plants.
Mountains, or a beach with some green around it. In a new city I go mural hunting.