Nikunj Bhavsar

Bridge part of the communication gap for Deaf and hard-of-hearing people by translating static ISL alphabet signs into text as they are made.

• Transfer learning on ResNet50 (ImageNet weights, frozen base) with a small custom classification head.
• Around 51,000 images: a public Kaggle ISL alphabet dataset plus roughly 800 self-collected images per letter.
• Augmentation (rotation, flips, zoom) and regularisation (L2, dropout, early stopping).
• An OpenCV pipeline that classifies webcam frames live.

About 96.4% test accuracy. The most common confusions were J and K, U and V, and C, I and L. Static signs only, and sensitive to lighting and background.

BSc final-year project, University of Hertfordshire

House prices alone say little about long-term investment potential. The question was which areas look strong once economic conditions, deprivation and safety are taken into account.

• Joined HM Land Registry Price Paid data, the ONS Postcode Directory, police.uk crime data and the Indices of Multiple Deprivation (2019) in PostgreSQL.
• Matched postcodes to boroughs with a nearest-neighbour search (SciPy cKDTree).
• Built a weighted investment score: low deprivation and high income raise it, high crime lowers it.
• Compared scores with price history from 1995 to 2023.

An interactive Streamlit dashboard with a postcode heatmap, score trends over time, and a price-against-score scatter for spotting areas that look undervalued.

Support early detection of diabetes from routine indicators: BMI, general, mental and physical health, age, education and income.

• Two-component PCA to explore the structure of the data.
• An RBF-kernel support vector machine on standardised features, tuned on a held-out validation split.
• Separate models for physically active and inactive groups, then each tested on the other group.

Roughly 73–75% test accuracy. The principal components showed substantial overlap between classes, which explains the ceiling.

Coursework, Machine Learning & Neural Computing, University of Hertfordshire

• Represented each track by tempo, timbre and a TF-IDF vector of its text description.
• Reduced dimensions with PCA and found nearest neighbours with KNN.

• Built in React Native, pulling places live from the Google Places API.

A 40% increase in user engagement time during testing.

• MVC architecture with a Java back end that logs every transaction.
• A Swing interface that updates in real time.

Handles more than 1,000 transactions.