I study what AI models hear and what they know: deepfake speech detection, recognising distress in spoken language, and how language models represent the world. I also build open-source tools for running LLMs and compete on Kaggle.
I'm a second-year Computer Science student at University College London, working where audio meets language models.
With Yupei Li and Prof. Björn W. Schuller, I co-first-authored work on efficient audio deepfake detection (IJCNN 2026) and on LLMs that recognise depression from speech (Frontiers in Computer Science, 2025). In summer 2026 I was a research intern at UCL with Prof. Lewis D. Griffin, probing how LLMs represent relationships between real-world entities.
Before research there was competitive programming (USACO Platinum in one season). Now I'm a Kaggle Competitions Expert with a silver and a bronze medal, and I like building things end to end: open-source tools for running LLMs, bug fixes for libraries such as OpenCV and Apache TVM, and hackathon prototypes.
I have more research in progress, including reinforcement learning and AI applications.
How models listen, and what they know.
How relationships between real-world entities are represented inside LLMs, measured with controlled fine-tuning and cross-domain evaluation, as a route to understanding the world models they acquire.
Bio-inspired dropin and neuroplasticity mechanisms that dynamically adjust the number of neurons in selected layers, making deepfake audio detectors more efficient and more accurate.
Injecting clinical psychological knowledge into LLMs, alongside Wav2Vec audio features, to recognise depression from spoken interviews.
Four papers, first or co-first author
* equal contribution
Where I've worked
UCL Department of Computer Science · supervised by Prof. Lewis D. Griffin
Built reproducible fine-tuning and evaluation pipelines across 313 real Wikipedia revision pairs, 36 country corpora, four model families and eight configurations, with controls that tested the metric itself. Across 777 dated facts, forced-choice accuracy beat open-ended generation by 0.29–0.39 in every model: the limit was eliciting knowledge, not missing it.
With Yupei Li & Prof. Björn W. Schuller · Remote
Audio AI, affective computing and NLP. Co-developed neuron-level dropin and plasticity for deepfake detection, and a knowledge-enhanced LLM pipeline for depression recognition.
University College London
Year 1 weighted average 70% (First-Class standard). Algorithms, Theory of Computation, Computer Systems, Object-Oriented Programming.
Things I've built
Tells you whether a vLLM deployment will start on your GPU, how much KV cache you'll get and how many requests fit. It reads a model's files from Hugging Face without downloading the weights.
Python · vLLM · pip install kvplan
Explains why a Linux disk is full, by cause: Docker, journals, caches, Hugging Face models and about 80 more, plus space that du can't see. It prints the fix and never deletes anything itself.
Python · C++ · Linux
Turns the outfit, food and places from your day into a playable pixel world that someone far away can walk around. It began as a four-hour hackathon project, built with teammates.
TypeScript · Next.js · OpenAI API · Canvas
A multi-agent network with live pipeline visualisation: inspect each agent's intermediate state and edit the execution graph at runtime. Non-English prompts slipped past filters that held in English.
Python · LangChain · AWS Bedrock · React
An agent network for the workplace: real-time transcription, answers in text or voice to match the question, and extracted knowledge saved to a Neo4j graph.
Python · Neo4j · LangChain
A desktop utility that makes sending emoji on WeChat fast. My most-starred open-source project.
C++ · Windows
Bug fixes I've sent upstream
Along the way
I'm looking for Summer 2027 internships in software engineering, ML research, research engineering or quantitative roles, and I'm always happy to talk about speech and audio AI, LLM tooling, Kaggle or competitive programming.