Ravi Mehta on building reasoning models that actually reason

Ravi Mehta has the kind of calm that makes you nervous. We met at a café in Shoreditch on a Tuesday morning; he ordered tea, not coffee, and started the conversation by asking what I'd written about reasoning models in the past six months. It wasn't a test exactly. It was the opening move of a person who builds for a living.

The forty-one-year-old founder of Lumen AI has been quietly assembling one of the most-watched AI companies in Europe. Last quarter, the London-based startup closed an $85 million Series B led by Sequoia, valuing the company at roughly $1.1 billion. It is, by most reasonable measures, a unicorn. Mehta finds the framing irritating.

"The valuation is the least interesting fact about the company," he tells me, twice, separated by about forty minutes. "What's interesting is what happens to enterprise software when you can finally trust the model to explain itself."

The thesis

Lumen's bet is straightforward to describe and difficult to execute: build models that don't just predict the next token, but produce a structured chain of reasoning that humans — and other systems — can audit, contest, and verify. The technical work is happening on three fronts: a custom training stack, a verification layer that runs on top of inference, and an enterprise control plane that lets compliance officers actually understand what a model did and why.

Mehta thinks the current wave of frontier models has solved the easy half of the problem. "Scale gave us models that produce plausible answers," he says. "We don't yet have models that produce defensible ones. That gap is the entire enterprise market."

Why now

The case he's making to customers — and there are now seventy-something Fortune 1000 names paying for Lumen's pilot — is that the next phase of enterprise AI adoption is going to be gated less by capability and more by accountability. Banks, hospitals, defense contractors, regulators: none of them care whether your model can write a sonnet. They care whether a procurement officer can later explain to a board why a decision was made.

"Compliance is the moat we keep underestimating," Mehta says. "Everyone in the Valley assumes the technical bar is what matters. It is — but only until the legal department gets involved. Then the bar moves." He gestures, faintly amused, at the irony: a former research scientist whose differentiation against rivals like OpenAI and Anthropic is increasingly about lawyers, not loss curves.

The bet

What's striking about Lumen's pitch is how unglamorous it is. There's no demo of an agent booking your flights. No video of a model passing the bar. What Mehta shows customers, instead, is a control panel: model decisions, the chain of reasoning that produced them, the data each step relied on, and an audit log that satisfies internal risk committees.

That austerity is intentional. "We're not selling AGI," he says. "We're selling deterministic-feeling AI to people who can't have surprises in their workflow." When pressed about how he competes against the frontier labs, his answer is sharper than expected: "We don't. We compete against the spreadsheet they're still using because they don't trust the model."

What comes next

The next twelve months will test the thesis. Lumen is shipping its general availability product in Q3 and has committed to a public benchmark for explainability — a deliberate provocation aimed at competitors. Mehta acknowledges the risk. "If we publish the benchmark and then a frontier lab matches us on it next month, we have a problem. But we also have a market that suddenly understands explainability matters."

I ask whether he's worried about being acquired before the bet pays off. He laughs, briefly. "I'd be a liar if I said the conversations don't happen. But the point of building this company was to find out whether reasoning models can become real infrastructure. You don't find that out by being absorbed."

The tea is cold. Outside, the morning has turned to a London drizzle. Mehta puts on his jacket and tells me he has to go — there's a customer call. Of course there is.

Lumen AI is hiring across research and infrastructure. For more coverage of the founders building the reasoning AI layer, see our latest interviews.