A selection of my data-science and machine-learning work. Code lives on GitHub and Kaggle.
Case Studies
Stille Ansprüche — German Social Benefits Discovery Assistant
RAG-based assistant that helps people discover German social benefits (Sozialleistungen) they may be entitled to but don't know exist — grounded in an official 500+ benefit inventory, with hybrid search and citations to the legal basis for every answer. Final project for the DataTalksClub LLM Zoomcamp.
Fraud Detection with Graph Neural Networks
Graph-based fraud-detection system that models financial transactions as an entity-interaction network (users, accounts, devices, merchants) and applies GNNs to detect fraud rings, collusion and anomalous behaviour.
Sentiment-Enhanced Stock Predictor
Combines historical stock-market data with news-sentiment analysis to improve price forecasting, integrating financial time-series features with NLP-derived sentiment signals.
Facial Emotion Recognition (CNN + AWS)
Recognises human emotions on the Kaggle Facial Expression Recognition dataset with a CNN, deployed as an AWS service via ECR, Lambda and API Gateway.
Satellite Image Segmentation (U-Net)
Trained a U-Net on an augmented set of satellite house images (25 of 30 roof-labelled), reaching 94% validation accuracy on T4 GPUs in Colab.
Medical Appointment No-show Prediction
Predicts whether patients show up for scheduled appointments to optimise resource allocation, handling imbalanced healthcare data.
Credit Risk Classifier
Built and deployed (through a REST API on a cloud service) a credit-risk classifier.
Marketing Mixed-Media Model
Building a Bayesian mixed-media model to interpret marketing performance.
Store Sales Forecasting
Probabilistic forecasting deep-learning model to predict store sales.
Soil Quality Monitoring
Deep-learning model to monitor soil quality and interpret results.
Business Intelligence Report
KPI-driven business intelligence report and revenue-optimization plan built from provided data.
Data Exploration & Visualization
Geospatial Mobility Analyses
Temporal and geo-spatial patterns in the Chicago taxi-trips dataset, developed as part of mobility-data work at Emmy.
Projects for Fun
Wordle Guesser
An algorithm that optimises guesses for Wordle using NLTK, Pandas, Matplotlib and regex.
