I build engineering systems that perform reliably in constrained environments — constrained by bandwidth, energy, budget, connectivity or data. The unifying thread is resilient sensing and intelligent infrastructure: acoustic underwater channels, low-power sensor networks, social housing testbeds, low-resource language models, and national climate-health platforms.

Project write-ups are on the projects page; papers and preprints are on the publications page.

Research Themes

Applied LLM & Agentic Systems

Retrieval-augmented generation and agentic workflows — chunking and retrieval evaluation, prompt and template management, guardrails and human-in-the-loop review. Fine-tuning open-weight models with QLoRA/PEFT for domain classification, and agentic AI on constrained hardware, with emphasis on failure modes, reliability and governance controls.

Low-Resource Language AI

Language technology for African languages, with PAWA AI, Brown University, and universities and schools in Tanzania:

  • PawaLID — a compact byte-level multi-task model for language identification across dozens of East African languages, beating the state of the art on identification and machine translation.
  • Handwritten Swahili recognition — custom OCR and vision-language models for handwritten Swahili academic documents, where frontier models fail.

Resilient Communications & the Internet of Underwater Things

Harsh acoustic channels, low bandwidth, high latency and severe energy constraints make the Internet of Underwater Things (IoUT) one of the most demanding areas in wireless engineering. My doctoral work introduced the DEKCS clustering and routing protocol (over 70% lifetime extension) and reinforcement-learning approaches to simultaneous wireless information and power transfer (SWIPT) using AUVs. Current work covers hierarchical federated anomaly detection (arXiv:2603.24648), semantic telemetry and visual-anchor navigation for AUV missions, and a survey of machine learning across the IoUT protocol stack (arXiv:2603.07413).

Smart Housing & Ambient Intelligence

Two IoT systems studies on an occupied social-housing testbed run with a UK social landlord:

  • Multi-sensor mould-risk forecasting — predicts risk 6–48 hours ahead (24-hour AUC-ROC 0.851–0.957, cross-home transfer up to 0.968).
  • Non-intrusive welfare monitoring for lone elderly residents — per-appliance power monitoring, environmental sensing, radiator valve states and circuit-level energy signals (F1 0.848 / 0.814; ~£200–300 per flat), without cameras, microphones or wearables.

Climate-Health Engineering & Open Statistical Infrastructure

Platform engineering that turns climate-health methods into institutional infrastructure. The climatehealth R package operationalises six indicators endorsed by the United Nations Statistical Commission. Through BREATHE (2025–2028) the work extends into causal attribution over a 100-million-person cohort.

Network Science for Public Resilience

Graph and network methods for mapping vulnerabilities and critical dependencies in UK supply systems, and machine learning for detecting entities using complex routing structures to breach sanctions.