Ugur Tuna · AI Leader & Systems Architect · London / Cambridge, UK
I build AI systems, and the controls that make them safe to use.
Technical AI leader bridging executive strategy, commercial ROI, and production systems engineering — spanning sovereign & agentic AI architecture, LLM evaluation, high-assurance governance, and regulated scale.
Leadership focus
Four areas where executive risk appetite, commercial economics, and engineering meet.
-
Sovereign AI architecture & TCO
Frontier vs. open-weights evaluation, private VPC & air-gapped deployment boundaries, tiered routing, and full unit-economics modelling.
-
Agentic controls & runtime safety
Policy-gated tool execution, execution-time re-validation, idempotency, tamper-evident audit logs, and kill switches for systems that act.
-
Evaluation & benchmarking
Automated eval harnesses measuring faithfulness, coverage, misstatement rates, human-judge calibration, and true verification cost.
-
High-assurance governance & data
Executive governance boards, policy-as-code assurance aligned to NIST AI RMF and OWASP LLM, plus national-scale regulated data delivery.
Perspective
- Now
- AI and technical governance lead setting organisation-wide AI strategy, risk appetite, and assurance boards with senior executives (Directors, C-suite, CISO, DPO) — while evaluating frontier and open-weights vendors, building commercial ROI models, and shipping hands-on LLM and agent workflow pilots across engineering, risk, and domain teams.
- Background
- Previously delivered production commercial ML (cutting forecasting error by 50% across 600,000+ SKUs for £700k in verified annual savings) and led a team of six data scientists at the University of Cambridge running national-scale biomedical and clinical data pipelines across 50,000+ participants.
- The thread
- I have operated systems where one wrong row reaches a patient study and models that move real commercial margin, and I now architect where autonomous AI meets enterprise accountability. Across frontier labs, regulated enterprises, and scientific platforms, the requirement is identical: evidence you can check.
Open-source proof
Four connected architectures answering one leadership question: how do you scale AI value without losing operational control?
-
Architect & price
Which deployment boundary meets sovereign rules, and what is the true monthly TCO once human checking and rework are priced?
ai-systems-architecture Python · Pydantic · 10 deterministic sovereignty & ROI gates · tiered routing simulator · ADRs
-
Control execution
How do you let an AI agent act without it sending the wrong thing, twice, to the wrong people?
agent-guardrails Python · SQLite approval queue · execution-time re-validation · hash-chained audit log · tool-use policy engine
-
Verify quality
Does an AI synthesis say what the sources say, what did it miss, and what did verification cost?
evidence-synthesis-eval Python · UK AISI Inspect · deterministic and model-graded scorers · blinded human-judge calibration
-
Govern & assure
How can a leadership team run proportionate AI assurance in an afternoon rather than a quarter?
ai-assurance-kit Python · Pydantic · JSON Schema · Jinja2 · NIST AI RMF & UK standards crosswalk
All systems & reference architectures, including clinical AI and genomics
How I lead
-
Activity is not impact.
Usage counts show exposure, not whether outputs were accurate or decisions improved.
-
Price total human effort.
Verification, review, and downstream rework belong inside the true ROI of every AI workflow.
-
A second model is a critic, not a verifier.
Models share blind spots; agreement between models is not ground truth.
-
Permission to build is not permission to deploy.
Connecting data, exposing tools, and authorising autonomous action are distinct risk gates.
-
Solve the data problem before adding an agent.
Deterministic automation or cleaner structure beats fragile autonomy whenever possible.
-
Surface uncertainty explicitly.
Systems that expose what is unverified earn trust; systems that hide gaps fail silently.
Scale & delivery
Selected delivery track record across enterprise AI leadership, commercial ML, and the University of Cambridge:
- Led organisation-wide AI strategy, governance boards (Directors, C-suite, CISO, DPO), vendor & model evaluations, ROI cases, and hands-on workflow pilots
- Delivered production demand forecasting across 600,000+ SKUs, cutting error by 50% for £700k in verified annual savings
- Led a team of 6 data scientists delivering national-scale biomedical and genomic operations
- Processed 50,000+ samples across imputation batches with automated strand-conflict resolution
- Delivered cryptographically verified research cohorts of 10,000+ samples and hundreds of GB into trusted research environments
- Harmonised multi-hospital clinical feeds toward the OMOP common data model across 10+ disease areas
- Executed regulated subject-access disclosures returning clinical genomic data for patient care decisions
Writing
Short essays on AI measurement, evaluation, and runtime control.
-
Usage is not quality: what to measure when you roll out AI
Usage counts show who clicked. They do not show whether the work got better, or what it cost to check.
-
A second model is a critic, not a verifier
Asking one model to review another can improve a draft. It cannot tell you the draft is true.
-
Read freely, write carefully: controls for AI agents that act
An agent that reads can mislead you. An agent that writes or sends can do something you cannot take back.
Connect
- GitHub
- github.com/dsugurtuna
- Writing
- dsugurtuna.github.io/writing/