Ugur Tuna: AI systems, and the controls that make them safe to use

In brief

Now: AI and digital transformation lead in the UK public sector. I lead an organisation-wide AI and automation programme: the governance and assurance that let people use AI with confidence, workflow pilots built hands-on with technical colleagues, and the skills staff need to use AI well.
Before: At the University of Cambridge, working on the NIHR BioResource, I ran end-to-end genomic data provisioning for one of the UK’s largest national research cohorts: HLA imputation across 50,000+ samples, recall-by-genotype studies, secure data releases and GDPR subject access requests.
The thread: I have built data systems where one wrong row reaches a patient study, and I now work where AI meets public accountability. In both, I care about the same thing: evidence you can check.

Three connected projects, one question: how do you get real value from AI without losing control of it?

Act safely

How do you let an AI agent act without it sending the wrong thing, twice, to the wrong people?

agent-guardrails Python · SQLite · hash-chained audit log · Claude tool use
Measure quality

Does an AI summary of the evidence say what the sources say? What did it miss, and what did checking it cost?

evidence-synthesis-eval Python · Inspect · model-graded and deterministic scorers
Assure and decide

How can a team do proportionate AI assurance in an afternoon rather than a quarter?

ai-assurance-kit Python · Pydantic · JSON Schema · Jinja2

Usage is not quality.

Adoption figures show who clicked, not whether the output was right.
A second model is a critic, not a verifier.

Models share blind spots, so their agreement is not evidence.
Permission to build is not permission to deploy.

Connecting data, sharing a tool and letting it act are separate decisions.
Fix the data before adding an agent.

Often a better template or ordinary automation is the right answer.
Count all the effort.

Checking and correction time are part of the cost of AI.
Show what is uncertain.

Say what has not been verified, every time.

At the University of Cambridge (NIHR BioResource):

Processed 50,000+ samples across HLA imputation batches, with automated strand-conflict resolution
Delivered genomic cohorts of 10,000+ samples with cryptographic verification
Ran secure cloud transfers of hundreds of gigabytes into trusted research environments
Harmonised clinical data from several NHS Trust formats into the OMOP common data model
Led a team of 6 junior data scientists in genomic data processing and secure data handling
Enabled research across 10+ disease areas, including rare diseases, IBD, neurodegeneration and immunology
Handled GDPR subject access requests that returned clinical genomic data for patient care decisions

Short, practical essays on using AI well.

Usage is not quality: what to measure when you roll out AI

Adoption figures 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.