Insights

From the frontier
of safety-critical AI.

Research, analysis, and hard-won perspective from a team building AI in environments where it is most difficult — and most consequential — to get right.

12 articles

Featured

Research
Safety-critical AI

Why safety-critical AI demands a fundamentally different engineering philosophy

The assumptions that make AI effective in consumer contexts are precisely the assumptions that make it dangerous in high-stakes environments. We examine why — and what a genuinely safety-first approach actually looks like in practice.

AK
Dr Amara Kone
March 2026  ·  12 min read
Read article →
Nuclear & Energy
The data problem in nuclear AI: why your sensor data isn't as useful as you think
Nuclear plants generate enormous volumes of sensor data. But sensor coverage, calibration history, and the relationship between readings and actual plant state are far more complex than most AI practitioners assume.
JR
James Reeve
Feb 2026 · 9 min
Regulatory
EU AI Act implementation: what high-risk operators actually need to do
The EU AI Act is in force. Most commentary focuses on what it says. This piece focuses on what regulated-sector operators need to actually do — and by when.
PO
Dr Priya Okonkwo
Feb 2026 · 11 min
Healthcare
Clinical validation of AI systems: what the literature gets wrong
The dominant validation framework — hold-out test sets, AUC scores, clinician comparison — systematically misses the failure modes that matter most in deployment.
SB
Dr Sophia Brennan
Jan 2026 · 14 min
Research
Uncertainty quantification is not optional in safety-critical AI
A model that cannot express how uncertain it is about its own outputs is not safe to deploy where those outputs will be acted upon. We explain what good UQ looks like and why most systems lack it.
SB
Dr Sophia Brennan
Jan 2026 · 16 min
Opinion
How regulated organisations should procure AI — and why most don't
Government and regulated-sector AI procurement processes are systematically selecting the wrong vendors and solutions. Here is what a better framework looks like.
AK
Dr Amara Kone
Dec 2025 · 8 min
Research
Domain-adaptive pre-training for regulated-sector NLP: our approach and early results
General-purpose language models perform poorly on regulated-sector text. We describe our DAPT approach across nuclear, healthcare, and financial domains, and share preliminary benchmark results.
LH
Lena Hoffmann
Dec 2025 · 18 min
Nuclear & Energy
Applying ALARP to AI systems in nuclear environments
ALARP is the foundational principle of UK nuclear safety. What does it mean to apply it to an AI system, and how should nuclear operators think about AI-related risk?
MT
Marcus Tate
Nov 2025 · 10 min
Security
Building ML systems for air-gapped environments: constraints and solutions
Most ML tooling assumes internet connectivity. Classified environments do not permit it. We explain what a genuinely air-gapped ML architecture looks like.
JR
James Reeve
Nov 2025 · 13 min
Regulatory
A practical guide to MHRA AI as a Medical Device classification
The MHRA AIaMD framework is the primary regulatory pathway for clinical AI in the UK. This guide covers classification, evidence requirements, and practical compliance steps.
PO
Dr Priya Okonkwo
Oct 2025 · 15 min
Opinion
The AI hype cycle is actively harming safety-critical adoption
Every inflated vendor claim in regulated sectors makes it harder for serious practitioners to have credible conversations with the organisations that most need AI done properly.
AK
Dr Amara Kone
Oct 2025 · 7 min
Healthcare
NHS information governance for AI projects: what you actually need
IG requirements for AI in the NHS are frequently misunderstood — both underestimated by vendors and overestimated by procurement teams. We explain what is actually required.
SB
Dr Sophia Brennan
Sep 2025 · 11 min

No articles in this category yet. More coming soon.

Stay informed

Browse by topic

Everything we publish, organised.

4+
Nuclear & Energy
ALARP, sensor data, ONR guidance, predictive maintenance
3+
Healthcare
Clinical validation, MHRA, NHS governance, NLP
3+
Research
Uncertainty quantification, DAPT, safety methodology
2+
Regulatory
EU AI Act, MHRA, FCA, ISO 42001

Want to work with the people
who write this?

The thinking on this page comes from the same team that builds the systems. If you are working on a problem in this space, we would like to hear from you.

Get in touch Join the team →