12 years of advanced AI R&D for the Ministry of Defence and FTSE 100 - from foundational research to modern LLM and agentic AI deployment. Oxford-based, internationally available.
I'm Jason Lee - Oxford-based Independent AI Consultant with 12 years building advanced AI systems for MOD, DSTL, and FTSE 100 organisations. I work at the algorithmic level: designing, leading, and delivering AI research that produces measurable outcomes.
I founded illumr Ltd in 2013, building it into a defence-accredited AI firm with international teams across the UK, France, Italy, and Bulgaria. In 2025 I transitioned to independent consulting, bringing a rare depth of experience directly to clients who need genuine AI expertise - not a resold off-the-shelf tool.
I sit at the intersection of deep AI research and strategic leadership. I work alongside brilliant researchers - often at PhD and post-doctoral level - directing advanced AI programmes, mentoring specialist teams, and translating what they deliver into clear value for boards and leadership. I understand the modern AI landscape well enough - LLMs, agentic systems, production deployment - to lead the people who build it and articulate its implications to those who fund it.
I now sit on the Steering Committee of the Cybersecurity & AI Governance Initiative (CAGI), helping close the governance gap between AI adoption and responsible deployment in critical infrastructure.
From literature review to proof-of-concept - at the algorithmic level. I design and lead advanced AI programmes across the full stack, from foundational research to production deployment.
Board-level AI strategy grounded in operational reality - not aspiration. Decision-ready roadmaps, governance frameworks, and ethics-led adoption planning that leadership can act on immediately.
Extensive track record within MOD frameworks including SERAPIS, DSTL, ARCD and DASA - with Research Worker Form clearance per project. Results presented at the MOD's AIFest event.
AI insight that cuts through the noise. Speaker at Buckingham Palace and the Milken Institute Global Conference, with international media coverage and published governance writing via CAGI.
The Ministry of Defence challenged us to demonstrate Generation-After-Next cybersecurity capability - 15+ years ahead of current. I wrote the bid, recruited an international team, and led a 3-phase programme training Reinforcement Learning Agents using Genetic Algorithms to autonomously defend against state-actor APT attacks.
Delivered a board-level AI strategy for the British Society for Antimicrobial Chemotherapy - mapping practical AI opportunities across accreditation, education, publications, and membership for a globally-distributed organisation. Decision-ready roadmap, governance framework, and board/staff communication pack.
Built a Fair Adversarial Network using GANs to eliminate bias from AI training data at source - not patching after the fact. 30× lower bias effect size in criminal recidivism forecasting models. Dockerised prototype deployable on private cloud. Selected for Accenture FinTech Innovation Lab London (1 of 5 in Data Stream).
Proprietary TDA engine uncovering hidden patterns missed by all other methodologies. PoCs with Pfizer, easyJet, Rolls Royce, Williams F1, Unilever, Banco Sabadell, HouseMark, Kier and more. Identified safeguarding clusters 10–50× more likely at risk for Metropolitan Housing. DASA-funded demonstrator for Defence & Security.
Strategy and governance work is most credible when grounded in live technical practice. AGTL is my personal R&D testbed — a portable intelligent guitar rack that puts modern embedded processors, edge AI, and agentic development into a single real-world build.
A rack-mounted DSP and AI platform built around an Electro-Smith Daisy Seed (ARM Cortex-M7 embedded DSP), an NVIDIA Jetson Orin NX (edge AI inference), and a multi-agent Claude-powered development workflow. Real-time audio effects run on the Daisy in C++; the Jetson hosts Python AI services for tone matching, LLM-driven parameter control, and speech-to-text interaction. The entire development process is agentic — built with Claude as an active collaborator using a custom multi-agent architecture.
Traditional cybersecurity frameworks were built for predictable systems. AI isn't predictable. Prompt injection, data poisoning, model inversion, agentic autonomy gone wrong - these require a fundamentally different approach. EN 304 223 is one of the first standards that has genuinely caught up.
Most discussions about AI risk start with the model - jailbreaks, hallucinations, poisoned training data. But some of the most serious failures begin much further upstream, compounding across data pipelines, deployment teams, governance structures and feedback loops before anyone notices. The missing discipline isn't model ownership or data ownership. It's system ownership - someone accountable for seeing the whole thing.
AI governance is entering its global phase - but fragmented, incompatible regulatory regimes create risk multipliers, not protection. From the EU AI Act to ETSI EN 304 223 to the UN Global Dialogue on AI, the argument is clear: the greatest risk isn't powerful AI. It's powerful AI deployed into weak, fragmented governance systems.
Available for contract engagements and select senior roles. If you're working on something serious in AI - defence, enterprise, governance, research, or agentic systems - let's talk.