Peer-Reviewed Research
Research output across graph neural networks, deep reinforcement learning for health, large-language-model evaluation, and quantitative methods in finance. Two papers accepted at ICHR-26, with additional manuscripts in preparation.
Graph Neural Networks · Spatiotemporal Modelling · Transport Network Resilience · Deep Reinforcement Learning · LLM Evaluation & Reliability · Retrieval-Augmented Generation · Agentic AI Systems · Cascading Failure Analysis · Self-Supervised Learning · AI for Health.
Interpretable spatiotemporal Graph Neural Network with integrated causal discovery, modelling relationships between climate variables and regional disease outbreak patterns across Pakistan. Validated against epidemiological ground truth. Awarded 1st Place with cash prize and Certificate of Excellence.
Soft Actor-Critic agent within a digital-twin simulation for autonomous insulin delivery, optimised for Ramadan fasting and high-glycaemic South Asian diets. Reduced post-Iftar hyperglycaemia by 35 mg/dL; zero severe hypoglycaemia versus 4.5% in standard protocols; outperformed clinical sliding-scale protocols in 92% of simulated scenarios.
Multi-dimensional statistical evaluation framework for benchmarking large language model factuality, consistency, and reliability across model architectures. Designed evaluation methodology and ran model-benchmarking experiments. Manuscript in preparation.
Quantitative asset-allocation algorithm using Bayesian optimisation to maximise risk-adjusted returns. Demonstrates robust out-of-sample performance with reduced overfitting versus traditional statistical backtesting baselines.
Designing robust asset-allocation algorithms that address overfitting in traditional statistical tests through regime-switching models and hybrid Bayesian / swarm-intelligence optimisation techniques.