Josh Le Grice

Normalised (3NF) relational database on Azure SQL for UK agricultural sustainability analysis, with automated Python ingestion pipelines and a production-ready REST API layer. Emphasis on data modelling, ingestion reliability, and scalable querying.

Coursework comparing full fine-tuning vs. LoRA on DistilBERT and implementing RAG for QA. Focus on efficient model adaptation and practical retrieval pipelines; trade-offs and deployment implications are documented.

End-to-end credit risk pipeline on real-world loan data. Compared Logistic Regression, Random Forest, and XGBoost; addressed multicollinearity and class imbalance. XGBoost achieved the best results with 92.8% Accuracy, 0.944 AUC, and 0.816 F1-score.

Implemented a full MLP in NumPy - forward pass, backpropagation, weight updates - achieving ~89% accuracy on Fashion MNIST, on par with a reference PyTorch implementation.

Identified top clinical predictors of diabetes diagnosis and benchmarked classification models. Logistic Regression yielded the highest accuracy at 77.3%, while SVM achieved the highest AUC at 0.850. Includes feature importance analysis and model interpretability.

Semi-automated reporting pipeline built during internship at Ekimetrics: raw data ingestion and cleaning through to interactive Power BI dashboards for commercial KPI tracking.

Streamlit web app for processing student data exports using a Bronze → Silver → Gold medallion architecture. Cleans and validates records, then splits into four exports for Power BI, membership, and survey platforms. Hosted on Azure App Service.