Classroom
to Company.

A comparison of four search algorithms (BFS, DFS, Dijkstra and A*) for evacuation routes on weighted grid maps, tested across 10 scenarios and 40 runs, with a Flask web interface to run them.

A* found optimal routes while exploring 68.6% fewer steps than BFS and 64.7% fewer than Dijkstra.

A stroke risk classifier (logistic regression) trained on a public dataset of 5,110 patient records, then tested against three fairness criteria: equal accuracy, demographic parity and equal opportunity.

The model caught 80% of stroke cases overall, but not equally. It caught 82.76% of cases in female patients and 76.19% in male patients, a gap that counts as bias under equal opportunity.

A Power BI solution on three months of supermarket till data (Q1 2019): data cleaned and audited, remodelled into a star schema with six dimensions, and DAX measures for sales, basket value and customer segments.

Every area of the marking rubric was rated Excellent to Outstanding.

Five question-answering systems, one without retrieval and four using different retrieval methods (BM25, TF-IDF, dense and hybrid), each answering the same 50 questions over a controlled test catalogue of 105 products, 10 retailers and 302 offers. Prices, stock and store details in the catalogue were synthetic.

Retrieval raised correct answers from 16% to 94% and cut unsupported claims by 85 to 90%. Dense retrieval performed best. Combining methods did not improve on it.

A comparison of three models (logistic regression, random forest and XGBoost) predicting heart attack history from the CDC’s 2022 public health survey: 442,067 adult responses and 66 engineered features. Missing data was handled by type, classes were balanced with SMOTE, and every model was tested on an untouched hold-out set.

Logistic regression caught 76% of heart attack cases (AUC 0.88). XGBoost scored the highest accuracy, 94.56%, but missed more cases. The study shows why accuracy alone misleads in screening.

A classifier sorting 3,064 public MRI images into three tumour types: glioma, meningioma and pituitary.

I built and trained both models, a custom CNN and a VGG16 transfer learning model, in TensorFlow and Keras. I also built the data pipeline (MATLAB files to images, a 70/15/15 split) and the evaluation code. Teammates led the literature review, poster and ethics analysis.

87.85% test accuracy on whole, uncropped scans. My report shows the gap to the published 99.2% came mainly from preprocessing, since that paper used segmented tumour images.

A Python application that fetches historical weather from the Open-Meteo API, stores it in a SQLite database of countries, cities and daily readings, and charts it with Matplotlib. Users can add new countries, cities and weather entries, with error handling and a written test plan.

Full marks for the database queries. The examiner found the interface not user friendly and the enhancements section weak (3 out of 10). Those are the first things to improve.