Alexandre
Berkovic

Somewhere right now, a compliance analyst is staring at an alert that means nothing. It is one of more than 100 million such alerts the world's rule-based systems will spit out this year. Most are noise. All must be read. Alexandre Berkovic decided that was insane - and worth a company.

That company is Sphinx, which he co-founded in 2024 and runs as CEO. Sphinx builds AI agents that do the unglamorous middle of financial compliance: the Know-Your-Customer, anti-money-laundering and Know-Your-Business checks that banks and fintechs grind through millions of times a year. The twist is where the agents live. Not in a new dashboard nobody wants to learn, but inside the browser - on top of the case-management tools, third-party portals, PDFs and email that compliance teams already stare at all day.

No engineering integration. No rip-and-replace. The agent just sits in the same screen the human does, reads the alert, runs the checks, drafts the request for information, and - the part regulators care about - writes down exactly why it decided what it decided.

First the music. Then the money.

Before any of this, Berkovic was in a very different business: sound. His first company, Adorno AI, built audio-generation models for video - the kind of multi-modal generative work he had been circling since his research days at Imperial College London and MIT. He co-founded it, ran it, and exited it.

A year into building audio tech, the founders looked up and went looking for a problem with more weight to it. They found one in a place nobody romanticizes: the back office of financial institutions. What they saw there was, in their own words, “ineffective and shockingly basic” - decades-old, rule-based software generating mountains of false positives while real fraud slipped through synthetic identities and deepfakes that a selfie and a utility bill can no longer catch.

So the audio company became a compliance company. The generative-AI instincts stayed. The riddle changed.

Sphinx's headline ambition, repeated by Berkovic and his investors, is to build “every financial institution's last compliance hire.” It is a provocative line, and a deliberate one. The bet is not that compliance teams vanish. It's that the grunt work does.

The arithmetic is the whole argument. When the overwhelming majority of alerts are false, the scarce resource isn't headcount - it's human attention. Move the repetitive review off people's plates and the same team covers far more ground, faster onboarding, fewer customers lost to verification limbo.

That is the gap the agents are sold to close. And it is why the “last hire” framing lands harder than a typical efficiency pitch. Compliance has long been the cost center every bank resents and none can cut. Berkovic is offering a third option: keep the people, lose the drudgery, and end up with a function that is both cheaper to run and easier to defend when a regulator comes knocking.

Whether the world agrees is the open question of his next few years. The seed round bought conviction and runway, not a verdict. But the shape of the bet is clear, and unusually legible for an AI company: take the most thankless, least visible work in finance, and make it move at the speed of the fraud it is meant to stop.

A short, fast line

Who wrote the checks

The Sphinx team is stacked deliberately - former compliance officers, PhDs and operators out of global banks. Berkovic's job is to keep the research instinct and the regulatory reality in the same room. Named customers include Equals Money.

There is a reason a generative-AI researcher ended up in compliance, and it isn't sentiment. Berkovic and his team make a sharp argument about timing: the tools banks use to prove you are who you say you are were designed for an earlier internet. A selfie. A utility bill. A document upload. Each of those was a reasonable proxy for identity right up until generative models learned to fabricate all three.

Synthetic identities - people who never existed, assembled from real and invented fragments - and deepfaked documents are no longer exotic. They are cheap. The same wave of AI that powers Sphinx's agents also powers the fraud Sphinx is built to catch. That symmetry is the whole point. Berkovic's view is that you cannot fight machine-speed deception with rule-based software written a decade ago. You need defenders that reason, adapt, and move at the same tempo as the attack.

Sphinx describes its larger ambition as building “the intelligence layer for global trust.” Strip away the grandeur and it is a practical claim: in a world where any document can be faked, the durable signal is the reasoning - a transparent, defensible chain of how a verification decision was actually reached. Not a verdict you have to take on faith, but a record you can audit.

A design engineer who learned to teach machines

It is easy to file Berkovic under “another YC founder.” The training underneath is less common. He studied Design Engineering at Imperial College London - a discipline that sits between making things and understanding why people use them - and then machine learning at MIT, where he worked on multi-modal generative AI, the branch that gets models to move fluently between sound, image and text.

That lineage shows up in the product. Sphinx is not a wrapper around a single prompt. It is a system of agents with assigned roles, designed to disagree before they conclude, built by someone who spent years on how generative models behave and where they fail. The design-engineering half explains the other obsession: meeting compliance officers exactly where they already work, inside the browser, rather than asking them to migrate their lives into yet another tool.

He keeps a public research profile and a personal portfolio site, quiet evidence that the academic thread never fully went away. The pivot from audio to anti-money-laundering was a change of subject, not a change of method. The method - point capable generative models at a stubborn real-world problem and engineer around their weaknesses - has been constant since the lab.