Machine Learning
Udacity Self-Driving Car Simulator
Trained a neural network to steer a simulated car from camera input — an end-to-end deep-learning pipeline from data collection to a model driving the track.
Deep learningComputer vision
Machine Learning · Recommenders
Hybrid Deep-Learning Course Recommender + Flask API
PyTorch hybrid recommender (sequence + content + collaborative signals) for 95,415 MOOC learners, served through a Flask REST API — beats item-CF on every metric.
PyTorchTransformerssentence-transformersFlask
Case study →
Predictive Maintenance
ReneWind: Turbine Failure Prediction with Neural Networks
Seven Keras architectures compared for early failure detection on ciphered wind-turbine sensor data — 93.6% recall on unseen failures.
TensorFlow/Kerasscikit-learnClass weighting
Case study →
AI & NLP · RAG
Medical RAG Assistant over the Merck Manual
Retrieval-augmented clinical Q&A with a local Mistral-7B over 4,114 pages — four-stage evaluation with documented failure analysis.
LangChainChromaDBMistral-7B (llama-cpp)PyMuPDF
Case study →
Machine Learning · Data Engineering
Insurance Claim Data-QA & Validation Pipeline
Leakage-safe pipeline that flags invalid German vehicle-repair invoices: automated quality checks, time-based validation, calibrated thresholds.
scikit-learnimbalanced-learnCalibration
Case study →Forecasting
Electricity Demand Forecasting: LightGBM vs XGBoost vs LSTM
Half-hourly grid-demand forecasting with lag, cyclic and decomposition features — LightGBM wins at 0.21% MAPE on validation.
LightGBMXGBoostLSTMFeature engineering
Case study →
Healthcare Analytics
Mental-Health Functioning (GAF) Statistical Analysis
Statistical study of 10,000 psychiatric patient records: hypothesis tests, feature screening, and an OLS model explaining 63% of GAF variance.
statsmodelsscipy.statspandasplotly
Case study →Big Data
Twitter Conversation Toxicity — PySpark EDA
Spark-based exploration of 28,823 Twitter conversations: cleaning, derived toxicity flags, and matched DataFrame-API vs SparkSQL analyses.
PySparkSparkSQL
Case study →Deep Learning · Experiment
LUNA25 Lung-CT Nodule Classification (3D CNN) — Experiment
Work-in-progress 3D CNN pipeline for lung-CT nodule classification: parallel dataset assembly from Zenodo, SimpleITK volume loading, 128³ inputs.
TensorFlow/KerasSimpleITKMultiprocessing
Case study →Healthcare Analytics
Heart Failure: Patient Characteristics & Mortality
Explored clinical records to identify which patient characteristics correlate with mortality risk, turning raw hospital data into interpretable risk factors.
PythonEDA
Data Analysis
Road Traffic Accidents: A Data-Driven Safety Study
Analyzed accident records to surface the conditions and locations where crashes cluster — evidence a safety program can act on.
Pythonpandas
Predictive Modeling
Real-Time Accident Analysis in London
Built models on live London accident data to predict severity and link outcomes to weather conditions as they change.
PythonModeling