Meet the engineer-CEO building the accuracy layer for fraud, AML, and identity — TFN

The popularity arrived earlier than the corporate turned two. Within the spring of 2026, the agency Aspire named Saarth Shah considered one of its 30 world founders to observe, a nod much less to the scale of his startup than to the issue he had chosen to deal with.

Shah, who was born in India and studied information science on the College of California, Berkeley after transferring from the College of California, San Diego, has spent most of his quick profession circling a single query. How do you make a pc’s reply about the actual world reliable sufficient to behave on?

He took it critically as an undergraduate. On the San Diego Supercomputer Heart, he constructed geospatial fashions that learn irregular 911 name patterns to alert authorities to mass-casualty occasions akin to college shootings. The stakes taught him early what accuracy prices. “One fallacious alert, and police and college authorities mobilise for one thing that seems to be nothing,” he mentioned. At Stanford’s Snyder Lab, the genetics group run by Michael Snyder, he helped construct Wearipedia, an open-source toolkit that makes information from wearable well being sensors usable for college students and smaller labs. The 2 efforts shared little besides the self-discipline beneath them, deciding how far a dataset may be trusted earlier than somebody depends on it.

He was constructing corporations the entire time. As an undergraduate, he co-founded Socale, a campus networking app backed by Berkeley SkyDeck and Blackstone LaunchPad that used a graph-based advice engine to match college students primarily based on educational pursuits. It handed a pair thousand downloads earlier than he moved on. In 2023, he co-founded Dart, the place he constructed a conversational voice AI receptionist for medical clinics on prime of GPT-3.5, Twilio, and Deepgram, a system that answered calls and booked appointments by telephone at roughly one-second latency and ran reside in two San Diego clinics. Between them, he took engineering jobs at Deepgram, the speech-AI firm, and at Whatnot, the live-shopping market, earlier than beginning Sixtyfour and becoming a member of Y Combinator’s 2025 class.

Sixtyfour builds analysis brokers that examine people and corporations for groups working in fraud, anti-money laundering, id, and due diligence. In an impartial analysis by a distinguished AI analysis lab, with outcomes set to be printed inside the subsequent week, Sixtyfour ranks as probably the most correct system of its type, forward of instruments from a lot bigger labs. Shah’s standing rests much less on the corporate’s dimension, which is modest, than on a place he argued early and constantly.

“Breadth is an arrogance metric,” he mentioned. “In fraud and id, the one quantity that issues is how typically you might be proper, and most AI analysis instruments are fallacious typically sufficient to be unusable. One fallacious reply and somebody will get banned, fined, and even convicted.”

He first made the case to an engineer at an identical firm, one which sells folks information to gross sales groups. You may be fallacious a 3rd of the time in gross sales and nonetheless work off stale data, Shah informed him, as a result of the price of a foul row is a wasted electronic mail. In fraud and id instances, the identical error charge ends in somebody being banned or arrested. It’s an retro factor to say in a market that has spent three years promoting protection, and his argument is that none of it counts if the reply is fallacious when it issues.

What separates Shah from many founders promoting AI is that he constructed the core of the product himself. He had chained language fashions with reside internet search right into a recursive loop, one which finds a chunk of proof, makes use of it to search for extra, and retains going till it has the complete image, constructed particularly to research folks. He knew it labored in January 2025, weeks earlier than OpenAI launched its personal Deep Analysis, when the loop related his personal on-line profiles to a childhood pseudonym he had as soon as used, a second digital id most instruments would by no means tie again to him. He architected that engine and the analysis techniques that measure its accuracy, the equipment that lets the corporate declare and defend its place on the general public benchmark. In an trade the place founders more and more narrate expertise another person wrote, his authority is first-hand.

That conviction shapes what the corporate builds and whom it sells to. Sixtyfour’s clients cluster the place a fallacious reply is dear, amongst blockchain-intelligence companies, funds corporations, and the trust-and-safety groups at on-line platforms, quite than the gross sales departments most information distributors court docket. It’s a smaller market than contact information by way of the variety of clients, however larger in monetary affect and extra demanding.

He runs a workforce of twelve, eight of them engineers, and has employed past the same old founder community. Amongst his early recruits is a trust-and-safety chief with 20 years within the discipline, previously the pinnacle of belief and security at Coupang, the South Korean e-commerce big. It’s a telling rent for a corporation that more and more sells to the individuals who police on-line platforms.

Colleagues describe a founder unusually allergic to the unverified declare, making use of to his personal product the usual Sixtyfour sells its purchasers: a function is judged much less by the way it seems in a demo than by whether or not its accuracy may be measured and defended.

Shah’s bigger declare is that the self-discipline he works in is about to turn into everybody’s downside. As AI-generated fakes flood marketplaces and overwhelm verification techniques, the platforms that run them are discovering that good-enough accuracy is now not ok.

“Each market is changing into a trust-and-safety firm whether or not it needed to or not,” Shah mentioned. “One fallacious name by our brokers and somebody will get punished, defrauded, or worse, arrested. Those that survive would be the ones that may inform, shortly and appropriately, who they’re truly coping with.”

The proof that he was early is beginning to accumulate. Marketplaces are reclassifying fraud as a belief downside that runs by way of the entire platform. Regulators are writing guidelines that assume real-time controls. The variety of corporations that must know, shortly and appropriately, who’s on the opposite facet of a transaction is rising sooner than the instruments that may inform them. Job marketplaces, rental platforms, and crypto exchanges that when outsourced belief to a vendor’s static file are actually constructing verification into the core of their operations.

The popularity and the income have arrived collectively. The Aspire nod barely registered, he says: the objective is to maintain constructing one thing folks need and use. He has made a behavior of arriving early to questions that later flip pressing, from emergency calls to scientific sensors to the id of a stranger on the far facet of a transaction. The sphere he has been quietly defining is beginning to catch as much as him.