Research

Research statement and directions

My research interests span multimodal and generative machine learning and their application to problems where data is complex and resources are limited. Three themes recur: models that learn from more than one modality, such as images, volumes, speech and text; large language models that can be trained, adapted and deployed efficiently, including for low-resource languages such as Yoruba; and learning systems that respect the structure of the scientific or clinical domain they serve.

These interests grew out of concrete work: generating radiology reports from 3D chest CT, inverting a physical radiative transfer model for Earth observation, and building real-time speech-to-speech systems for an embodied humanoid. Each exposed the same questions from a different angle: how to represent heterogeneous inputs, how to keep models faithful and interpretable, and how to make them work under real constraints of compute, data and time.

The directions below describe the broad areas I want to work in rather than a fixed programme. Where a direction connects to work I have already done, the entry says so; where it points to work I intend to do, it says that too.

Multimodal Machine Learning and Vision-Language Models

How can models learn jointly from images, volumes, speech and text, and ground what they generate in what they perceive?

Most of the data that matters in the real world does not arrive as a clean 2D image or a short string of text. I am interested in models that combine modalities, including volumetric and spectral data, audio and language, and in generation methods whose outputs can be checked against their inputs rather than merely sounding plausible. That includes the practical side: how such systems behave when they run in real time and interact with people.

Themes

Vision–language and speech–language models
Cross-modal representation learning
Multimodal generation and evaluation
Embodied agents that perceive, reason and act

In my work

Radiology report generation from 3D chest CT (Springer LNICST, PAAISS 2025), and real-time speech–language systems for an embodied humanoid at IQAI.

Generative AI and Large Language Models

How can large generative models be trained, adapted and deployed efficiently, extended to low-resource languages, and used reliably as reasoning components in larger systems?

Large language models are increasingly used as components: planners, operators, tutors, translators of intent. I want to understand when that is sound, how to make it affordable, and how to extend it to languages and settings that current models serve poorly. That spans efficient training and deployment under compute constraints, retrieval and grounding, and evaluation that reflects how a model is actually used rather than how it scores on a leaderboard.

Themes

Efficient fine-tuning, quantisation and on-device deployment
Multilingual and low-resource LLMs
Retrieval-augmented generation
LLMs as planners and operators in complex systems

In my work

A NeurIPS 2025 AI4NextG workshop paper on LLMs as operators of 6G radio access networks; LLM-driven speech systems at IQAI; building NLP resources for Yoruba.

Medical Imaging and AI for Healthcare

How can AI make medical imaging and clinical documentation more accurate and more accessible, particularly where specialist expertise and compute are scarce?

Clinical imaging is where complex data, expert language and real consequences meet. I am interested in models that work with full volumes rather than isolated slices, that justify their findings and quantify uncertainty, and that remain trainable and deployable in settings without large computing budgets, which is often where the need is greatest.

Themes

Volumetric imaging (CT, MRI)
Report generation and clinical language
Data- and compute-efficient training
Clinically meaningful evaluation

In my work

My MSc thesis and first peer-reviewed paper generated radiology reports from CT-RATE volumes on a single-GPU budget.

AI for Science and Scientific Machine Learning

How should learned models and scientific models be combined so that predictions respect known physics and transfer from simulation to real observations?

Scientific problems come with forward models, conservation laws and instruments whose physics are understood. Treating these as inductive biases rather than obstacles yields models that are more data-efficient and more interpretable. I am interested in this combination wherever a trustworthy simulator exists, from vegetation and climate monitoring to the physical and biomedical sciences.

Themes

Physics-informed and simulation-based learning
Inverse problems and parameter retrieval
Earth observation and environmental monitoring
Scientific foundation models

In my work

Physics-informed inversion of the PROSAIL radiative transfer model on Sentinel-2 imagery during my InstaDeep fellowship (arXiv:2511.10387).

Structured and Interpretable Machine Learning

How can we build models whose representations and uncertainties can be understood, inspected and trusted?

Interpretability, for me, is largely a question of structure: if a model’s latent variables are the parameters of a physical process, or the findings a clinician would list, then its reasoning can be read, audited and corrected. I am interested in architectures, training objectives and evaluation practices that make this the default rather than a post-hoc explanation.

Themes

Structured and physically meaningful latent variables
Uncertainty quantification and calibration
Interpretable architectures
Evaluation beyond benchmark scores

In my work

The PROSAIL inversion work produces posteriors over physical parameters rather than opaque embeddings, with prediction intervals checked against field measurements.

Reasoning and Representation Learning

What representations and training signals let a model reason over heterogeneous inputs and act on them under real-world constraints?

The thread connecting my work is the ambition to build systems that do more than recognise patterns: systems that combine evidence across modalities, respect known structure, and act within limits of time and compute. Volumetric imaging and scientific inversion gave me the problems; production speech and robotics systems gave me the constraints.

Themes

Representation learning across modalities
Reasoning in language and multimodal models
Agents and embodied AI
Learning and inference under compute constraints

In my work

Real-time multimodal systems and robot control for an embodied humanoid at IQAI, and the formal framework of the LLM-RAN operator paper.