Hi, I'm Dinuja Perera

Built a transformer-based NLP system to analyse employee reviews from 500 UK companies. Compared models including BERT, RoBERTa, and XLNet, achieving 76% accuracy. Integrated topic modelling and aspect-based sentiment analysis to extract key insights from text data.

Developed a machine learning solution to identify customers likely to subscribe to term deposit products. Performed exploratory data analysis and feature engineering on customer and campaign data. Applied K-Means clustering to segment customers based on behaviour. Used DuckDB and SQL queries within Python to analyse structured datasets efficiently. Helped identify high-value customer groups and reduce unnecessary marketing effort.

Built a machine learning solution to detect dissatisfied customers in logistics delivery services using operational and feedback data. Identified patterns indicating poor delivery experiences and predicted likely dissatisfaction. Used LazyPredict to benchmark multiple classification algorithms and Hyperopt to optimize hyperparameters for improved performance.

Applied supervised machine learning to predict the severity of road accidents across the UK using 2019 public data. Evaluated Random Forest, SVM, Decision Tree, KNN, and Deep Neural Networks. The deep neural network achieved the highest accuracy of 80.65% in classifying accidents as 'Slight,' 'Serious,' or 'Fatal.'

Investigated metal part manufacturing datasets to predict part lifespan and classify defects. Regression models (Linear, Lasso, Ridge, Random Forest) were compared, with Random Forest achieving 97% accuracy. For defect detection, both binary classifiers and CNNs were tested. K-Means clustering revealed distinct process parameter groups influencing part quality.

Developed causal machine learning models to understand employee turnover behaviours. Used Random Forest and statistical feature importance techniques to identify the most influential drivers of churn. Performed feature leakage detection, refined model input space, and enhanced model generalisation. Integrated email sentiment analysis as an additional behavioural signal. Work carried out in a Linux-based environment using object-oriented Python.

Explored transformer-based approaches for detecting hate speech on YouTube and Reddit using the ETHOS dataset. Compared BART and RoBERTa for binary and multi-class classification. BART achieved 70% F1-score and 58% top-1 accuracy, outperforming RoBERTa in distinguishing hate categories including gender, race, and religion.

Developed machine learning approaches to analyse gas turbine sensor data and detect early signs of system faults. Used correlation and autocorrelation analysis to distinguish normal and faulty sensor behaviour. Applied time-series modelling techniques to support predictive maintenance strategies.

Tested 12 deep learning architectures, including RNNs, CNNs, transformer-based models (e.g., BERT, RoBERTa), and hybrid architectures (e.g., CNN + LSTM) to detect hate speech on social media platforms.