Research
My research asks how multimodal models can be adapted efficiently, how they can reason over diverse evidence, and how AI agents can act reliably as their tasks and environments change.
Research programme
01
I develop parameter-efficient and semantically grounded methods that repurpose foundation models for new domains and tasks. This work spans visual prompting, machine unlearning, retrieval-guided adaptation, and lightweight model updates.
02
I study how models combine visual observations, language, and structured knowledge to reach grounded conclusions. This includes document-understanding systems that must trace answers to evidence and reason consistently across related forms.
03
I investigate agents that learn from interaction, adapt to changing tasks, expose evidence for their decisions, and preserve useful prior capabilities. The goal is dependable behaviour across the full lifetime of an agent.
Publications
EMNLP 2026 · Industry Track
A semantic-first framework for generating privacy-preserving enterprise form packets with ontology-grounded extraction, question answering, and validation signals.
AAAI Conference on Artificial Intelligence · Oral
A retain-data-free method for targeted unlearning in vision-language models, using hyperbolic geometry to remove concepts while preserving surrounding knowledge.
Waste Management
Prompt-guided segmentation for construction and demolition waste, supporting box, point, and text interaction in cluttered environments.
AAAI Conference on Artificial Intelligence
A parameter-efficient way to adapt vision-language models by grounding visual prompts in language, including black-box settings.
Laboratory Investigation
A two-stage computer-vision system for detecting and classifying white blood cells in bone-marrow aspirate images.