Research That Ships.

One of the first agentic AI systems built for the legal domain — built before MCP (Model Context Protocol) servers existed and ahead of the industry's broad adoption of LLM-based agentic approaches. The system moved employment-law and industrial-relations service delivery from a manual, task-level process to an autonomous, multi-step AI reasoning system.

Outcome: Secured an 18-month first-to-market advantage for the client by delivering agentic AI capability well ahead of mainstream adoption.

Proposal focus (verbatim from the grant record): "Research a fully automated employment law and industrial relations service enabled by Artificial Intelligence."

An AI-driven Open Source Intelligence (OSINT) platform combining NLP and image analysis to accelerate and automate the collection of due-diligence intelligence on individuals and entities — used for Know-Your-Customer (KYC) and Anti-Money-Laundering (AML) risk screening and third-party/vendor-risk workflows, for a client base of regulated financial institutions (banks, insurers, wealth managers) and private equity firms.

Scope of ownership: Directed a cross-functional, international AI initiative spanning a 30+ person organisation with headquarters in the UK and India. Partnered directly with the company Founder and two regional managers; owned budget allocation (staff vs. compute) and the 1–3 year strategic technology roadmap.

• Production model achieving 98% classification accuracy
• Reduced analyst annotation time by 30%, and effort by almost 98%
• Tracked post-deployment revenue impact and ran A/B testing with real users
• Collaborated with legal/compliance teams to secure joint University-Corporate IP and execute patent filings

AI-driven models to optimise customer interactions, deployed directly into BT's internal production systems. A unified transformer-based multi-task model for joint intent classification and named-entity recognition was developed as part of this work and is deployed and actively used by BT's development team.

Outcome: Reduced annotation time by 90%. The work scaled the effectiveness of BT's full Data Science team (50+ staff). Owned the strategic commercial roadmap deciding "what gets built and why." Resulted in two research papers, a production deployment, and the KTP Associate securing a permanent role at BT — the project concluded ten months early as a direct result. Collaborated cross-functionally with BT's Legal and university compliance teams on regulatory standards and joint patent filings.

An integrated smart platform and bespoke client-side application (site plan material sharing, management and version control) for the property development industry, improving office-site communication. The AI agent — the first of its kind for this use case at the company — reduced construction workers' on-site decision-making time. Engineering build implemented in .NET by the supervised KTP associate/engineer, integrating into Mersea Homes' existing systems.

Outcome: Saved the company an estimated £50,000/year. Winner, "Best KTP Project (SME) 2022."

Personally architected a commercial generative AI product end to end — fine-tuned and deployed small language models (SLMs) that captured personalised, local-community language patterns for users. Owned the full stack: architecture, fine-tuning, evaluation, and Kubernetes/Docker production deployment. Unified previously disconnected GenAI efforts into a single delivery engine.

Full project name: PICASO — Probabilistic Conceptual Spaces Sensemaking under Uncertainty. Spans the Ministry of Defence (end customer), the Alan Turing Institute, and the University of Southampton.

A production-ready intelligence platform for the Ministry of Defence, built heavily on generative AI (LLM-based reasoning). As grant-holding Principal Investigator, directed the R&D roadmap for modelling knowledge as probability clouds, achieving state-of-the-art accuracy with built-in uncertainty quantification for high-stakes decision-making. Outcomes are backed by peer-reviewed technical documentation.

Project page →

Directed teams to use Amazon AWS to develop RoleCatcher's CV search systems. Introduced several novel additions, including automatic cover-letter generation — built when LLMs were still in their infancy, well ahead of mainstream adoption.

Outcome: Aided the startup's growth and scaling. The collaboration was covered by multiple media outlets as an example of a successful university-industry partnership building a startup, and of research models applied directly to industrial production systems.

Media coverage: University of Essex News →

Joint research initiatives with Thomson Reuters Labs (London and India research labs) on Bayesian Kernel Uncertainty for domain-specific document modelling and reliable predictions, including reliable medical predictions. Published: "BAKER: Bayesian Kernel Uncertainty in Domain-Specific Document Modelling" (WSDM 2025) and "Would You Trust an AI Doctor? Building Reliable Medical Predictions with Kernel Dropout Uncertainty" (WISE 2024).

Recognition: Delivered invited talks to Thomson Reuters' London and India research labs (2023). Talk coverage →

A Knowledge Transfer Partnership project on advanced security and open-source intelligence (OSINT) capability.

Recognition: featured in press releases highlighting high-impact security solutions. Media coverage: University of Essex News →

R&D on exploiting VR technology and NLP for police training and analysis, under the proposal "Exploiting VR Technology and NLP for London Metropolitan Police."

NG-CrAI — a Next-Gen Crowd-AI framework for scientific discovery, combining crowdsourcing and AI to support scientific research workflows and scientometric analysis.

Designed a novel non-parametric temporal graphical model for mining time-aware urban living styles — introducing a new temporal dimension to MSRA's existing "LifeSpec" urban human-computing model, improving its time/space complexity, output quality, and handling of data sparsity. The proposal was independently authored, and the resulting model formed the basis of a production-scale urban-lifestyle analytics platform processing data from 1M+ individuals across 493 cities.

Designed, developed and deployed a CPU load-balancing architecture for a Linux cluster, and a multi-threaded intranet crawler.

Recognition: Awarded a Letter of Appreciation.

Advised on curriculum design and AI strategy for the Computer Science Engineering programme.