AI researcher working on LLM reasoning, mechanistic interpretability, and agentic systems.
I'm an AI Engineer at Technovative Solutions (Manchester) and a Research Fellow at Fatima Fellowship. My work sits at the intersection of multi-agent reasoning, verifiable evidence, and the tools we use to tell when a model is actually thinking — and when it's just confabulating.
I started out in undergrad trying to explain black-box rumour-detection models — why does this LSTM flag a post as a lie? That question has followed me through every project since, and has kept sharpening.
If a model's output can't be audited — traced back to evidence, to a rationale, to a circuit — then for high-stakes work it may as well be a coin flip with confidence.
My MSc pushed me toward graph neural networks: structure is a useful prior when you're trying to reason about relations between claims and evidence. At Fatima Fellowship I got to put that into practice — first as a cross-graph GNN, then as an agentic pipeline (MERIT) where LVLMs plan, call web-search tools, and build verifiable evidence graphs before committing to an answer.
What I want to do next, in a PhD, is go one level deeper: not just build reasoning systems but understand them. Mechanistic interpretability of multi-step reasoning, especially in agentic settings where one model delegates to tools or to other models. How does reasoning compose? Where does it break? Can we build agents whose chains we can audit the way we audit circuits?
Chain-of-thought is convenient but unfaithful. I'm interested in what counts as actual reasoning inside a forward pass, and how the scratchpad relates to the computation.
Circuits-level methods — SAEs, causal mediation, activation patching — applied to the reasoning traces I care about, not just toy tasks.
What happens when one reasoner delegates to another? Tool-calls as a window into compositional structure; inter-agent communication as a proxy for intermediate representation.
Outputs are only as good as the rationales they sit on. Structured evidence (graphs, citations, tool traces) is how I keep systems honest.
An agentic pipeline that orchestrates large vision-language models with web search to build verifiable evidence graphs and traceable rationales before committing to a verdict. The point isn't accuracy — it's that you can read the trace and check the work.
We formalise claims and evidence as separate graphs and score image-text consistency with a cross-graph attention GNN. The attention maps turn out to be interpretable; the accuracy (93.05%) beats the strongest LLM baseline by 2.82 points.
Applied explainable-AI techniques (LIME, attention attribution) to interpret black-box rumor detection models trained on social media data. The question that started everything: why did the model say that?
Open-source reference implementation of the MERIT agentic pipeline: LVLM planner, web-tool executors, and evidence-graph builder with a traceable verdict endpoint.
Wrapped IBM/NASA's Prithvi Earth-Observation foundation model as a containerised inference service with self-supervised vision backbones (SimCLR) for downstream climate-adaptation workflows.
PyTorch Geometric implementation of the cross-graph attention network behind the CVPR workshop paper. Includes attention-map visualisations that actually say something about the model's reasoning.
Notebooks on RAG, LoRA/QLoRA fine-tuning, RLHF/GRPO experiments, and a few interpretability sketches.
Python · PyTorch · PyTorch Geometric · LangChain · RAG · LoRA / QLoRA · RLHF / GRPO · Hugging Face · FastAPI · Docker · AWS (EC2/S3) · GCP Vertex AI · W&B.
If you're a professor whose lab aligns with these questions, or a team hiring for research-engineer work where the research is still real — I'd love to hear from you.
Usually respond within 24h.