AI Researcher & Engineer working on adaptive machine learning for the messy, drifting signals around and within real human bodies — biosignals, audio, and the systems that try to read, understand and generate them.
Most ML works when the world holds still. Real-world signals don't — so my work is about the models that keep up.
I'm an AI researcher with a PhD from the University of Edinburgh (January 2026), supervised by Prof. Kianoush Nazarpour. My thesis — Mitigating Confounding Factors in Myoelectric Control Through Adaptive Modelling and Learning — sits exactly where deep learning meets the inconvenient reality of biosignals: electrodes shift, postures change, users adapt to the model while the model is trying to adapt to them.
Before finishing the PhD I worked on multimodal sensor fusion and personalising foundation models at Meta Reality Labs in New York, building wristband interaction models that had to run with low latency in genuinely out-of-distribution conditions.
My research targets the core failure modes of deployed ML on continuous sensor data: non-stationarity, cross-user and cross-session distribution shift, label scarcity under real-world annotation budgets, and signal corruption from noise, motion artefacts, and environmental variability. These are shared bottlenecks across biosignal processing, wearable sensing, and audio — speaker variability, domain mismatch, acoustic noise robustness, and low-resource adaptation are structurally the same problems in a different modality. I currently work on extending these methods to audio and foundation-model adaptation. I work end-to-end: from experimental design and data curation through model development to deployment optimisation.
Currently, I'm looking for the next place to do this work in earnest — foundation models for audio and biosignals, Health AI, neural interfaces, or any research group that takes real-world robustness seriously.
Static benchmarks tell us about a frozen world. The models that ship are the ones that keep working in the second week.
Annotation budgets are the real constraint in clinical ML. Active learning and curation are first-class research, not engineering chores.
EMG, EEG, ECG, IMU — each has its own physics. Architectures that respect the signal generalise farther than those that ignore it.
The wearer adapts too. Co-adaptation is the design problem nobody warns you about — and the one most worth solving.
Audio LLMs are quietly failing at the thing they're supposed to do. Recent work shows that much of the gain from reasoning-tuned audio models comes from reasoning over text, not the audio, and that longer chains of thought often degrade performance as the model's grip on the audio fades through the chain. Reasoning, where it works, is unevenly useful: necessary for some prompts, wasted for others, and silently harmful when the model can't actually hear what it claims to hear.
Training-free decoding strategies that force the model to re-attend to the audio at fixed points in its reasoning chain, paired with a mechanistic analysis of where audio attention drops out.
A learned router that decides, per prompt, whether reasoning helps or hurts — treating inference compute as an acquisition function rather than a fixed budget.
Same framing as the biosignals work: distribution shift, adaptive compute, and the budget constraints that decide whether a model survives contact with deployment.
A research agenda for the post-scaling era of physiological ML. Below: the questions I want to spend the next decade on, organised from signal to system to person.
Biosignals drift across days, electrodes, postures, sweat, and mood. The interesting science is identifying which shifts matter and building models that don't quietly fail when reality moves.
PPG, EMG, ECG, EEG. What does pretraining on heterogeneous bodies actually give us, and what's the right tokenisation for a noisy, non-stationary, low-SNR signal?
From session-level recalibration to lifelong adaptation. Closing the loop between the model and the wearer without catastrophic forgetting or runaway drift.
When labelling is expensive — clinic time, expert annotation, RLHF — what is the smallest budget that still moves the model? Acquisition functions as first-class research objects.
EMG + IMU + audio + vision. Different noise floors, different sampling rates, different latencies. Fusion that respects the asymmetries rather than averaging them away.
Most biosignal benchmarks measure within-session accuracy. Deployment doesn't. I'm interested in OOD-by-design evaluation, not OOD as a post-hoc surprise.
Skin, muscle, electrode placement and impedance all have a physics. Architectures that encode it as inductive bias generalise farther than those that brute-force it from data.
The wearer is not a passive data source. They learn the model while the model learns them. Co-adaptation has its own dynamics and its own failure modes.
Translational work for prosthetic and rehabilitation users. The last 20% of distribution shift kills more devices than any other failure mode, and it lives in people's homes.
Low-power, low-latency inference on wristbands, patches, and implants. The constraints aren't an afterthought — they're often the most interesting part of the problem.
Designing evaluation protocols that surface failure modes before deployment does. Statistical rigour and cross-subject splits as a design principle, not a footnote.
The reason for all of the above. Augmenting human capability — clinical, assistive, expressive — rather than automating around it.
Position paper benchmarking domain generalisation methods for robust myoelectric control across varying arm positions and contexts in EMG-based control. Accepted at the EurIPS Workshop on the Science of Benchmarking and Evaluating AI.
Position paper on deep learning approaches for EMG-based gesture recognition with improved feature representations. Published in IEEE Transactions on Neural Systems and Rehabilitation Engineering.
End-to-end experimental study isolating arm translation — same gesture, same user, different limb position — as a distribution shift in EMG-based control. Open dataset and benchmark released for reproducible work on adaptive prosthetic control.
A human-in-the-loop active learning framework for long-term adaptation of myoelectric prostheses, defining optimal query strategies for robust EMG-based control.
Applied deep learning to in vivo calcium imaging data from freely-behaving mice, predicting spatial navigation and exploratory behaviour through neuron activation mapping inspired by class activation techniques.
Sequential deep learning approaches for EEG signal analysis, addressing time dependencies, non-stationarity, cross-channel connectivity, and memory mechanisms in brain–computer interfaces.
3D visualisation tool for MRI brain scans using Unity and Microsoft HoloLens, designed to improve spatial data comprehension for medical students. Awarded Best Student Project, Department of Computer Science, Swansea University.
End-to-end ML pipeline for continuous navigation tasks via wristband biosignals, including supervised fine-tuning of foundation models, data curation, and deployment optimisation. Real-time inference for continuous gesture recognition with reduced latency, adapting off-the-shelf models to novel biosignal tasks for next-generation AR/VR input.
An open EMG dataset and benchmark for one of the most under-studied confounders in myoelectric control: arm translation. Same gesture, same user, different limb posture in space — and the decoder falls apart. The dataset isolates that distribution shift, and the accompanying benchmark evaluates classical and deep learning approaches under it. Published in Nature Scientific Data. Repo hosted by the MoveR Digital Health & Care Hub.
Simulation fine-tuning Mistral-7B with an Active Preference Learning acquisition function for DPO, compared against random sampling under a fixed annotation budget. Win-rate plots show APL reaching higher performance with substantially less labelled preference data — same adaptive-data thesis from EMG, ported to language model alignment.
Volumetric MRI rendered as an interactive 3D mesh in Unity, deployed to Microsoft HoloLens so medical students can manipulate brain anatomy in their physical environment. Awarded Best Student Project, Department of Computer Science, Swansea University.
Thesis: Mitigating Confounding Factors in Myoelectric Control Through Adaptive Modelling and Learning. Successfully defended December 2025. Focus on distribution shift, active learning, and domain generalisation in EMG-based control of upper-limb prostheses.
Co-led a course bridging technical AI implementation with entrepreneurial problem-solving. Guided students in developing MVP solutions for real-world health and technology challenges.
Developed large-scale multimodal sensor fusion models for continuous navigation tasks in wristband-based interactions. End-to-end ML pipeline including data curation, SFT of foundation models, and deployment optimisation — contributing to next-generation AR/VR input systems.
Intensive ML training across modern deep learning, probabilistic methods, and large-scale systems.
Computer vision and deep learning applied to in vivo calcium imaging of freely-behaving mice (with Prof. Oisin Mac Aodha & the Centre for Discovery Brain Sciences) — neuron activation mapping for behaviour prediction.
Sequential deep learning for EEG analysis in decision-making contexts — time dependencies, non-stationarity, and cross-channel connectivity for brain–computer interfaces.
Object-Oriented Programming (Java), data structures, and algorithms — for first- and second-year undergraduates.
Final-year project: HoloLens for medical neuroimaging education (Best Student Project, College of Science Award).
Co-organised the WiML workshop, the largest community gathering for women researchers in machine learning, co-located with NeurIPS.
Co-organised a workshop bringing together clinicians, end-users, and ML researchers to scope adaptive, human-centred rehabilitation technology.
Workshop on participatory design for neuroprosthetic systems, foregrounding the lived experience of prosthesis users alongside technical research.
Co-taught a course on bridging technical AI implementation with MVP product development for health and technology contexts.
A practical collection of academic writing habits and go-to paper-writing resources.
A taxonomy of what breaks myoelectric control and why it's harder to fix than it looks.
A vision for local-first AI built on biology-inspired architectures.
The next role.
A collaboration.
A good problem.