Why “AI Leader
and Strategist”?
Because for the last decade I have been the person who decides what gets built, by whom, with whose money, and at what risk — across five organisations, from a Ministry of Defence intelligence platform to a GenAI start-up joined at its founding. My job title says Associate Professor. My job, in practice, has been to direct AI programmes. Below is the ledger.
The three objections.
If you doubt the title, it is almost certainly one of these. Each one is a reasonable challenge, so each one gets a direct answer.
In the UK, a funded Knowledge Transfer Partnership or Innovate UK Smart Grant makes the academic the accountable owner: you hold the budget, hire the engineer, set the roadmap, and answer to both a funding body and a paying company. I have held that seat four times, at Neotas, BT, Wyser and Mersea Homes.
Dual accountability, not adviceOn the funded engagements, largely no — and that was the point. The build sat with the full-time research engineers I recruited, funded and directed, embedded inside the partner company. My deliverables were the technical roadmap, the budget split between staff and compute, the risk calls, and the hiring decisions. That is the leadership work; it is what the money was for.
Direction was the deliverableA 98%-accuracy classification model in regulated financial due diligence; a transformer model live inside BT’s production systems cutting annotation time 90%; one of the first agentic legal-AI systems, 18 months ahead of the market; a defence knowledge-graph platform shipped open source with a live analyst GUI. See the engagements →
Production, not prototypesThree UK terms, translated.
UK research funding has its own vocabulary. If you hire in industry, here is what these actually mean on an org chart.
A full-time, salaried engineer — usually MSc or PhD level — recruited onto the project and embedded inside the partner company’s own team for the duration, typically two to three years. Not a student, not an intern, not a placement. I ran the recruitment, held their budget line, and set their technical direction; the company gave them a desk and a production codebase. Several were hired permanently by the partner at the end.
A Knowledge Transfer Partnership: a government co-funded scheme (Innovate UK) where a company part-pays for a multi-year R&D programme run inside its business. The company is a paying customer with a commercial stake, so delivery is judged on what shipped — and there is a funding body auditing it in parallel.
Principal / Co-Investigator is a grant-holding designation, not a job title. The PI is the person the money is awarded to and the person answerable if it fails: they own the budget, the hiring, the roadmap and the risk register. The nearest industry equivalent is a Programme Director or Head of AI R&D for that workstream.
Why I stayed in academia.
Not because industry didn’t call. Because the academic seat is the only one that let me run AI R&D across many companies at once, on horizons no quarterly roadmap would fund.
The bet horizon
My doctoral thesis argued that word order and segment structure materially improve text models — years before transformers proved it at scale. No product roadmap funds a five-year bet. A research programme does.
The access
In one decade: defence (MoD), telecoms (BT), legal (Wyser), regulated finance (Neotas), retail (eBay), property (Mersea), policing (Met Police), and government policy. No single employer gives you all eight at once.
The mechanism
The UK deliberately built Knowledge Transfer Partnerships and Smart Grants so that academics hold real industrial budgets and hire their own engineers. That mechanism is why my leadership record is £1.87M across twelve awards rather than a line on someone else’s org chart.
The compounding asset
Products depreciate; people don’t. 17 doctoral researchers and four embedded engineers now sit at BT, Huawei Research, Lloyds, 3 UK and IIT Gujarat. Building the talent supply is leverage no single build gives you.
The adversarial audit
Every claim on this site survived anonymous peer review at NeurIPS, AAAI, ACL and SIGIR. In industry, a bad AI decision hides behind an NDA for years. In academia it is torn apart in eight weeks — which makes for better decisions.
And it wasn’t a fallback
I held Director of Engineering & Chief Scientist at Bezoku concurrently with the academic post, owning the entire technical direction of a GenAI start-up from its founding stage. Staying was a choice, made while doing the other thing.
One operating model, eleven organisations.
The funding label changes. The seat never does: I define the problem, hold the budget, hire the builders, own the risk, and answer to the customer.
AI Leadership.
Strip the job titles away and leadership is a short list of verbs: fund, hire, direct, de-risk, govern, hand over. Pick any one — here is the named, linked evidence.
- Neotas — £350k, Principal Investigator. Owned the end-to-end P&L, the split of budget between staff and compute, and the 1–3 year technology roadmap for the AI function of a 30+ person company across the UK and India — partnering directly with the Founder and two regional managers. Engagement detail →
- Wyser — £400k Innovate UK Smart Grant, PI. My largest single award; I wrote the proposal and directed the build of an agentic legal-reasoning system to market. Detail →
- PICASO — £140k Alan Turing Institute, PI. Set technical direction and budget priorities across three institutions simultaneously. Project deep-dive →
- I write the proposals that create the budgets. On the £230k travel-safety programme I authored the complete work package and delivery plan; on the Microsoft Research Asia award I designed the model, wrote the technical proposal, and presented it to the Microsoft judging panel myself.
- Non-cash resourcing too: won a dedicated NVIDIA Academic Hardware Grant (Quadro RTX6000) to resource the multi-modal programme's compute — because a roadmap without GPUs is a wish.
- Directly recruited and hired full-time research engineers (the KTP Associate role — see the translation above) across four industry partnerships, each embedded inside the partner company for two to three years. I ran the selection, the onboarding into the business, and the day-to-day technical direction thereafter.
- Scaled effectiveness across teams I didn’t own. At BT, the models I directed cut annotation time 90% across a 50+ person data science organisation. Leadership there meant changing how a large team worked, with no authority over its org chart. Detail →
- 17 doctoral researchers supervised, including a Doctoral College Director’s Award 2026 winner (Ubaid Azam) and the first Portuguese PhD student ever awarded a Microsoft Research PhD Fellowship (António Correia). Where they are now →
- Placement is the KPI. Alumni are at BT, Huawei Research, Lloyds Bank and 3 UK, and one is now an Assistant Professor at an IIT in Gujarat. On the BT programme, the research engineer I recruited was hired permanently by BT on the strength of the delivered work — which ended the project ten months early.
- Built an organisation from nothing, twice over: founded Competitive Coding at Southampton, secured dedicated funding from the Head of School, and led student teams to international competition.
- A £230k programme was withdrawn mid-flight when the partner company took COVID-era financial losses. I don’t hide it: running a portfolio means some positions close early, and the discipline is protecting the people and the salvageable work when they do.
- Regulatory risk, handled with the regulators’ own teams. On both Neotas and BT I worked directly with corporate Legal and Compliance functions on regulatory standards, joint University–Corporate IP, and executed patent filings — in KYC/AML and telecoms, two of the least forgiving compliance environments in the UK.
- I build risk into the models themselves. PROCS, which I designed for the Ministry of Defence, replaces point estimates with calibrated probability so the system can flag when not to trust its own output — 45.1% vs 26.3% Hits@1 on high- versus low-confidence predictions. For a defence analyst, knowing the machine’s uncertainty is the deliverable.
- Data-governance risk by design. FLiP (LREC 2026) lets partners train collaboratively with zero raw data crossing organisational boundaries — 16% trainable parameters, 90% less GPU memory, compatible with differential-privacy regimes. That is an architecture chosen to satisfy an information-governance constraint, not a benchmark.
- Validation against reality, not test sets. Neotas ran live A/B testing with real users and tracked post-deployment revenue impact. Cross-cultural work was reviewed by questionnaire panels across India, Japan, China, Italy and the Middle East.
- PICASO spans the Ministry of Defence (the end customer), the Alan Turing Institute, and the University of Southampton — delivered under the Turing’s Defence & Security Programme and a GCHQ framework agreement. I set technical direction and budget priorities across all three. See PICASO →
- Shipped, not shelved: a public GitHub repository, a pip-installable
pypicasopackage, and a live analyst GUI. A defence knowledge graph of 68,826 entities and 143,532 relations across 355 relation types was built to support it. - Internationally distributed delivery at Neotas across UK and India headquarters, and a research programme evaluated across five countries.
- Set direction inside someone else’s research lab: co-designed the experimental methodology and technical direction with a scientist at eBay Research America; eBay’s global leadership selected the work as a Top 3 Leaders’ Choice Award winner out of 2,000+ submissions. CIKM 2021 paper →
- Cross-institution research leadership also with Thomson Reuters Labs (London and India), INESC TEC (Portugal), and Microsoft Research Asia.
- Royal Society Pairing Scheme 2022. One of a small national cohort of scientists selected to be paired with a Whitehall civil servant (Home Office) to inform government policy-making on science and AI. Announcement →
- United Nations Development Programme. Selected through a global open competition to lead a team on Re-imagining Trust and Safety for AI. UNDP programme →
- Associate Editor, AI Communications; Guest Editor, IEEE Transactions on Computational Social Systems; Distinguished Review Board, ACM Transactions on the Web.
- Senior Programme Committee for AAAI (2024–26), ICASSP 2025, ECAI (2023–25), IJCNN 2025 and ACL Rolling Review — the seat that decides what counts as credible AI work at the field’s top venues.
- Curriculum and institutional strategy: invited member of the Board of Studies for Computer Science at Adani University, India; Visiting Fellow at the University of Essex; Deputy Appeals Officer and PhD progress panel member at Southampton.
- I wrote the whole thing, then handed over the lead. On the £230k automated travel-safety-alerts programme I authored the complete work package and delivery plan and was set to lead as PI. I stepped back to Co-PI so that a colleague could hold their first PI role. Developing another leader was worth more than another line on my CV.
- Systems outlive the funding. Every production system — BT’s internal environment, Neotas’ classifier, Mersea’s platform, Wyser’s reasoning engine — kept running after its grant closed.
- Handovers to the public, not just successors: PICASO shipped open source with a package and GUI so analysts can use it without me in the room.
- Covering when others can’t — taking over lectures, tutorials, final-year project supervision and the entire undergraduate admissions portfolio at short notice when colleagues left or went on leave. Unglamorous, and exactly the job.
Every pound, named.
Twelve competitively won awards. Six as Principal Investigator, six as Co-Investigator or Co-PI. Nothing here is a share of someone else’s grant.
Plus internationally denominated awards not counted above: HKD 692,894 (Hong Kong RGC General Research Fund, Co-I) and USD 20,000 (Microsoft Research Asia Urban Informatics Fund, Co-I), an NVIDIA Academic Hardware Grant (in-kind compute), and a three-year BT–Essex doctoral scholarship held as PI and main supervisor. Full grant record in the CV →
AI Strategist.
Leadership is executing well. Strategy is choosing the right thing to execute before the market agrees with you. The only honest test is a track record of early calls that later turned out right.
At the time, mainstream text models threw word order away. My doctoral thesis proved that modelling word order and segment structure measurably improves text models, against the prevailing bag-of-words consensus.
✓ Vindicated by the sequence-aware attention in every transformer and LLM sinceWhile the field standardised on point vectors, I argued that words should occupy regions and densities in space to capture vagueness and hierarchy — matching Skip-gram on standard tasks while beating unsupervised baselines on hypernym detection (CoNLL 2017), and later outperforming GloVe outright (AAAI 2019).
✓ The same geometric idea now underpins my MoD uncertainty workDirected the RoleCatcher CV-matching build for a start-up and added automatic cover-letter generation — while LLMs were still in their infancy. The company grew on it and the university–industry story was picked up by national media. Detail →
✓ Now a standard feature of every hiring platformCo-designed the methodology and technical direction for a multi-modal hypergraph click-through model with eBay Research America, fusing visual and structural signals when most production ranking was still tabular. CIKM 2021
✓ Top 3 of 2,000+ eBay submissions · Best Paper nominee · ACS Digital Disruptors Award winnerWon £400k and directed one of the first agentic AI reasoning systems for the legal domain — built before Model Context Protocol existed and well ahead of mainstream agentic adoption. Detail →
✓ Delivered an 18-month first-to-market advantageWhile the market optimises benchmark scores, I bet the MoD programme on calibrated uncertainty and selective prediction: a system that declines to answer is deployable in defence; a confident, wrong one is not. See PICASO →
● Live bet — shipped open source, judge it yourselfDirected a four-year, three-grant multi-modal programme on cultural fit in generated content — starting in 2021, long before cultural alignment became a mainstream evaluation concern. Programme page →
✓ Culminated in CALM at NeurIPS 2025, as senior authorFive decisions, and what they cost.
Calling the direction is only half of it. The other half is picking one option and living with the trade-off — which is what these five actually were.
Abandoned a working classical-vision pipeline for diffusion models.
Phase 1 of the cultural programme worked (Mask R-CNN, DeepFace, OpenPose). It was also a dead end for quality. I retired it and moved Phase 2 to Stable Diffusion with CLIP embeddings and cross-attention control. Sunk cost is not a strategy.
Chose closed-form geometry over sampling-based Bayesian inference.
For PROCS I picked diagonal Gaussians with O(1) inference via direct geometric formulas instead of slower sampling methods — matching computationally heavier models like NBFNet while adding uncertainty estimates. Chosen for deployability, not elegance.
Designed for information governance first, accuracy second.
Defence and finance partners cannot pool raw data — so FLiP moves only compact prompt vectors across boundaries: 16% trainable parameters, 90% less GPU memory, differential-privacy compatible. The constraint became the architecture.
Ran eight sectors in parallel instead of specialising in one.
Defence, telecoms, legal, finance, retail, property, policing, publishing. It costs depth in any single vertical. It buys pattern recognition across all of them — and it is why five separate organisations could be taken from zero AI capability to production.
Gave away the PI role on a £230k grant I had written.
I authored the full work package, then stepped back to Co-PI to let a colleague hold their first programme-owner role. Reads as a loss on paper. It was capacity building — and honest about my own load across other funded programmes.
Chained three grants into one programme rather than chasing three papers.
GCRF, then an NVIDIA hardware grant, then the Royal Society Pairing Scheme — all funding one continuous research line over four years, evaluated across five countries. That sequencing is what produced a NeurIPS paper instead of three disconnected outputs.
Other people’s judgement, not mine.
Every item below was awarded by a selection panel, a leadership team, or a competitive global call.
Check the primary sources.
Every claim on this page traces to a grant record, a peer-reviewed paper, a public repository or a third-party announcement. Here is where to look.