Two former Google DeepMind researchers have quietly launched Sampura Research, a startup that wants to make AI systems easier to audit and trust. The company has raised $11 million in early funding.
The problem they're chasing
Sampura's pitch is straightforward: AI models are getting more capable, but the tools to check whether they're behaving as intended haven't kept pace. The founders argue that evaluation — figuring out if a model is safe, unbiased, and actually doing what it's supposed to do — is the weak link in the AI pipeline.
Their answer is something they call hybrid AI oversight. That means combining automated checks with human judgment in a way that's more rigorous than either approach alone. The goal, they say, is to build systems that don't just perform well on benchmarks but can be trusted in real-world deployment.
What the money goes toward
The $11 million isn't a giant round by AI startup standards, but it's enough to hire a small team and start building. The company hasn't named any clients or public pilot programs yet. What they have said is that the funding will go toward developing tools that help organizations stress-test their AI before it reaches users.
The founders are both alumni of Google DeepMind, the lab that produced AlphaGo and large language models that power many commercial products. They've declined to say why they left, but the move fits a pattern of DeepMind veterans starting their own safety-focused ventures.
Why evaluation is suddenly a big deal
AI evaluation is a growing concern across the industry. As models are used in hiring, healthcare, and finance, the consequences of a silent failure get more serious. Sampura's approach is to treat evaluation as a continuous process rather than a one-time test — the hybrid part means human reviewers can step in when automated metrics miss context.
The company says its work addresses critical evaluation challenges, but they haven't detailed exactly which challenges. That's typical for a startup at this stage. What's clear is that they see a gap between how AI is built and how it's checked.
The startup is still early. No product launch date has been announced, and the team is likely hiring. The real test will be whether their hybrid oversight model can scale beyond research papers and into actual deployment.
For now, the $11 million gets them a runway. The next milestone will be showing that their tools work on a real system — and convincing someone to pay for it.
