I'm Alexander Pugeat

Built an end-to-end analytics platform that mines 200k+ high-elo Teamfight Tactics matches to surface meta team compositions via unsupervised ML, identifying 35+ distinct archetypes — including several untracked by sites like MetaTFT. Engineered a Python ingestion pipeline against the Riot Games API into a normalised PostgreSQL schema (Dockerised, psycopg2 ETL), with daily automated refreshes via launchd. Clustered compositions using MultiLabelBinarizer features, UMAP (Jaccard) for dimensionality reduction, and HDBSCAN. Frontend in TypeScript/Next.js with Tailwind, deployed on Vercel.

Built an end-to-end ML pipeline in TensorFlow/Keras to detect, classify, and denoise simulated supernova neutrino events in 10,000+ Liquid Argon Time-Projection Chamber (LArTPC) detector images. Engineered physics-based noise simulation, a ResNet-18 classifier robust to heavy noise (98% accuracy), a residual U-Net denoiser, and a Bayesian-optimised regression model that cut energy-prediction error by 45% (to 8% MAE).

Built a Python/MATLAB pipeline to extract signal from noisy, high-throughput photon-count time series, developing a novel tail-fitting methodology for multi-exponential decay models. Applied weighted least-squares regression with custom preprocessing (statistical truncation, residual filtering) to handle heteroscedastic noise, validated against ground truth at 98.0% accuracy on lifetime, 97.5% on amplitude, and 96.3% on rotational correlation times.

Pioneered a novel application of shaking tables (a mining technology) for shoreline microplastics removal. Reduced a 6-dimensional parameter space to 2 key drivers using Plackett-Burman fractional factorial screening and ANOVA significance testing. Fit quadratic response-surface models via Central Composite Design on a synthetic-sediment dataset, optimising recovery to 88.1%, separation efficiency to 75.3%, and purity to 85.8%. Diagnosed distribution shift on out-of-sample River Thames field data (recovery dropped to ~60%) and traced the gap to unmodelled covariates (organic matter, biofouling) altering feature distributions.

Built a Python simulation pipeline (NumPy, Matplotlib, PyWake) modelling stochastic wind patterns via Weibull/Rayleigh distributions and vectorised power-curve calculations to forecast turbine output, validated against the DTU PyWake benchmark — optimising site and turbine selection to a Levelised Cost of Electricity of £42/MWh (14% below market) at a 63.4% capacity factor (408 GWh/year).