Vishal Yadav.
"Making AI more honest about its own failures."
I'm a researcher and engineer based in London. I think a lot about what it means to build AI responsibly - what it takes to measure, audit, and understand the systems we're deploying at scale. Occasionally I surface for air to read, cook, or argue about films.
This is the less formal version of me. · The CV is over here →
How I got here.
Chapter one
Indore → London
I grew up in Indore, India, and studied computer science at RGPV. Research found me through an internship at IIT Indore - I was building AR applications and working on autonomous vehicles, and somewhere in the middle of that I realised that building things that think was what I actually wanted to do.
I moved to London for an MSc in AI (Computer Vision & Robotics) at Queen Mary University of London. Graduated with Distinction. It was during the MSc that I started working on clinical AI - and things got more interesting.
Chapter two
When the stakes became real
I spent two years as a Research Assistant at QMUL working on bias in mental health electronic health records. The question we were asking was simple: do AI diagnostic models treat patients differently based on who they are?
The answer was yes - and the mechanism was surprisingly mundane. Something as ordinary as note length correlating with patient gender was enough to propagate bias into diagnostic outcomes. That work was published in Nature in 2026. It made the stakes of AI safety feel very concrete - not theoretical, not distant.
Chapter three
Building the infrastructure
Now I build evaluation frameworks and observability infrastructure for deployed AI systems. The clinical work taught me that AI failures are often detectable - if you know what to measure and where to look. That's the problem I'm working on.
Trustworthy AI needs to be measurably trustworthy. Not just in papers, but in deployment - in the places where it actually touches people's lives.
What's occupying my mind.
Evaluation as a discipline
How do we actually know if an AI system is doing what we think it's doing? Benchmarks are often measuring the wrong thing - optimised for legibility, not for what actually matters. I'm interested in evaluation not as a checkbox but as a genuine epistemological problem that the field hasn't fully reckoned with.
Bias in clinical AI
My research showed that something as simple as note length correlates with patient gender, and that this propagates into diagnostic models. It made me think harder about what "data quality" really means when the data is human-generated language full of unstated assumptions - written by people, about people, in a world that is not neutral.
What AI safety actually requires
There's a lot of theoretical work in alignment, and a lot of empirical work in red-teaming. I'm interested in the gap between them - specifically, what observability and evaluation infrastructure you need to detect when a deployed system is drifting from intent. Safety needs to be operational, not just conceptual.
The limits of explainability
I've used LIME, GradCAM, and integrated gradients in production research. I'm genuinely uncertain whether current explainability techniques tell us what we think they do. The gap between "here's which pixels mattered" and "here's why the model decided this" is still enormous - and closing it matters.
Building things that last
Software engineering and research have very different cultures around reliability and maintenance. I find myself thinking about how you build AI systems that are auditable, maintainable, and honest about their own limitations - not just at launch, but over months and years of deployment.
Books I recommend.
The Alignment Problem
Brian Christian
The clearest account I've read of why AI safety is genuinely hard - and why the people working on it are not being paranoid.
Sapiens: A Brief History of Humankind
Yuval Noah Harari
Changed how I think about collective belief, the stories we tell ourselves, and what actually drives human behaviour at scale.
Homo Deus
Yuval Noah Harari
Where Sapiens explains where we came from, this asks where we're going. Unsettling in the best way - especially if you work in AI.
Bhagavad Gita
Ancient text
On duty, action without attachment to outcome, and staying grounded when things are uncertain. I come back to it often.
1984
George Orwell
More relevant every year. A reminder that the architecture of systems - not just their intent - determines what they do to people.
12 Rules for Life
Jordan Peterson
Disagree with parts of it, but the core argument - that meaning comes from taking on responsibility - has stuck with me.
Atomic Habits
James Clear
Less about motivation, more about systems. The insight that you don't rise to your goals, you fall to your systems, is one I think about regularly.
Outside of the lab.
Reading
Philosophy of mind, history of science, and fiction that plays with structure or perspective. I read slowly and reread often - the Bhagavad Gita being the clearest example.
Cooking
Mostly South Asian and Mediterranean. I treat recipes as starting points rather than instructions - the interesting part is always the improvisation.
Films
I take them seriously and have strong opinions I'm always happy to argue about. Ask me about a film and you'll find out quickly whether we agree.
Community
Helped organise Isohack, a 36-hour hackathon. Mentored Master's students through their dissertations. Care about making technical spaces more welcoming - in practice, not just in intention.
Languages
I grew up thinking in Hindi, work in English, and I'm still not entirely sure which one I dream in.
Find me.
Say hello.
I'm best reached by email. Open to conversations about AI safety, research collaborations, or anything else that seems worth discussing.