Research & innovation
My research bridges academia and industry - from predictive healthcare AI at Queen's University Belfast to production AI systems deployed at scale. Ph.D. in Artificial Intelligence & Data Science.
What the work is about
Production AI systems
A model that only works in a notebook has not been tested yet.
The questions I find interesting start where the demo ends: what happens on a four-year-old phone, on a gallery floor, with the network down, when the input is nothing like the training set. Most of my systems work is about the distance between a result and a system that survives strangers using it.
- On-device computer vision for museum visitors — camera processed entirely on the device, only abstract presence events leave it
- Battery drain cut by more than 50% on low-end devices, which decided the whole inference architecture
- A jurisdiction-based retrieval engine for legal contracts, where the applicable law is a first-class retrieval dimension rather than a filter
AI and product engineering
Enterprise AI mostly fails after the model works, not before.
Roughly 95% of enterprise AI projects produce no measurable effect on profit and loss, and it is almost never the modelling that breaks. It is the data path, the failure modes nobody specified, the integration nobody owned, the operational reality a demo never has to survive. This is the subject of the book, and the reason I wrote one rather than another paper.
- Rule-based RAG prompt engines — dynamic prompt construction from definitions, templates and composable rules
- Evaluation and verification as first-class engineering, not a phase at the end
- A low-code framework that cut development cost by roughly 30% across an engineering organisation
Applied AI research
Decisions under uncertainty, and languages the field tends to skip.
My doctoral work at Queen's University Belfast looked at how AI-driven decisions hold up inside real healthcare systems — predicting length of stay, and the discretisation methods that pipeline depends on. Alongside it runs a longer thread of work on Sinhala, a low-resource language with very little published NLP behind it, and on generative audio.
- PhD, Artificial Intelligence & Data Science — Queen's University Belfast, 2020–2024
- Two novel methods: a discretisation approach using mixed exponential models, and Progressive Coxian phase-type dynamic Markov models
- Early Sinhala-language NLP for hate-speech detection, and an APICTA Gold Award for generative music
Publications
Machine learning model classifying Sinhala songs into distinct genres from audio features including rhythm, pitch, and timbre - supporting music recommendation and cultural preservation. Supervised research with S.M.K.N. Nawamini Karunathilake.
Predictive Modelling for Length of Stay with the MIMIC-III Critical Care Database
Healthcare AI application predicting patient length of stay using the MIMIC-III critical care database, enabling data-driven resource allocation in clinical settings. Joint work with Prof. Adele Marshall and Dr. Aleksandar Novakovic at Queen's University Belfast.
Using Mixed Exponentials for Unsupervised Discretization
Novel approach to unsupervised data discretization using mixed exponential distributions - improving preprocessing quality for downstream machine learning pipelines in clinical and industrial datasets.
Komposer - Automated Musical Note Generation Based on Lyrics with Recurrent Neural Networks
AI system for automated music composition that generates melodic note sequences from text inputs using recurrent neural networks. The underlying system (Komposer) was awarded the APICTA (Asia-Pacific ICT Alliance) Gold Award in the Tertiary category.
Identifying Racist Social Media Comments in Sinhala Language Using Text Analytics Models with Machine Learning
Natural language processing model for detecting racist and hate-speech content in the Sinhala language - one of the first low-resource language hate-speech detection systems using text analytics and supervised machine learning.
Forecasting Monthly Ad Revenue from Blogs Using Machine Learning
Machine learning approach for predicting monthly advertising revenue from blog platforms - exploring regression models and feature engineering on web analytics and content metadata.
Virtual Airplay Drum Kit Based on Hand Gesture Recognition
Computer vision system enabling real-time virtual drumming through hand gesture recognition - combining image processing, gesture classification, and audio synthesis for an interactive music performance experience.
A Simple Machine Learning Approach for Identifying Promotional Short Message Service (SMS) Messages
Text classification model for distinguishing promotional SMS messages from personal messages, using machine learning with feature engineering on message content and metadata.
Interested in collaboration?
Open to research partnerships, speaking engagements, and consulting on AI system design and deployment.