§The quiet floor on New Montgomery Street.
On any given Tuesday, somewhere between your morning coffee and your second meeting, three million Americans try to open a bank account, a credit card, a buy-now-pay-later plan, or a car loan. SentiLink's API gets a look at most of them. It does not ask for selfies. It does not demand a passport. It looks at the application and decides, in milliseconds, whether the person on the other end actually exists.
That is the entire business. It is also, somehow, an $85 million company.
SentiLink occupies a corner of San Francisco's New Montgomery Street that looks more accountancy than tech. There is no neon, no espresso bar trend piece, no founder podcast. The hallways are quiet because half the staff are statisticians who would rather be left alone with a confusion matrix. The other half are risk operations analysts who, by trade, do not get excited about much.
Above: a sentence the founders would probably underline, then quietly delete.
§A crime nobody named for a decade.
For most of the 2010s, American banks had a fraud category they did not know how to label. Applications would come in - real-looking Social Security numbers, real-looking names, real-looking addresses - get approved, fund, then vanish. The customer would never call. The customer, it turned out, had never existed.
The industry called it bad debt. The actual term was synthetic identity fraud: a Frankenstein record stitched from a child's unused SSN, a fictional name, and a plausible address. It was, depending on whose numbers you trust, a $20-billion-a-year hole in the U.S. financial system. Nobody had a product for it. The credit bureaus could score risk on real people; they had nothing useful to say about people who were not.
This is the gap SentiLink walked into. Not a new technology, exactly. A new category. Which is harder.
§Two risk people, one borrowed blessing.
Naftali Harris was Affirm's first data scientist. He arrived with a Stanford master's in statistics and the kind of unflashy temperament the role rewards - more interested in being right than in being heard. Max Blumenfeld ran risk operations alongside him. Between them they had seen the underwriting funnel from both ends, and they had seen synthetic identities slip through it more times than they cared to admit.
In 2017 they pitched their boss, Affirm CEO Max Levchin, a polite version of "we are leaving to build the thing that should have caught those." Levchin's response was famously short and famously generous. He wrote a check.
Andreessen Horowitz wrote a bigger one in 2019 - $14 million, Series A. Two years later, Craft Ventures led a $70 million Series B that brought the total to $85 million. Felicis and NYCA came along. Nobody in the round mistook SentiLink for a moonshot. The pitch was, and is, deeply unromantic: there is a measurable hole in the financial system, and we will charge per look.
A short, mostly accurate history
A timeline written by people who hate timelines.
§Scores, signals, and a phone call.
What SentiLink actually sells, technically, is a handful of API endpoints. The Synthetic Fraud Score answers one question: how likely is this applicant not a real person? The ID Theft Score answers the cousin question: how likely is this real person not the one filling out the form? Facets is the wholesale version - raw identity attributes lenders can fold into their own underwriting math.
Underneath, the work is less elegant. The models eat consortium data from hundreds of institutions, public records, SSA verification, behavioral patterns, and a slowly-curated archive of confirmed-fraudster fingerprints. The closest analogue, oddly, is a credit bureau - except the bureau is grading you on whether you exist, not on whether you pay.
There is also a human floor. SentiLink staffs a team of risk analysts who manually review the edge cases - the ones the model flagged at .51 - and call lenders back with a verdict. In a category where false positives cost real customers real loans, the phone call is the part nobody else wants to do.
Funding, in slabs
A chart for people who don't usually read charts.
§The customer list reads like a checking account.
SentiLink is not a household name. SentiLink's customers are. The roster runs through the top U.S. banks, regional credit unions, fintechs, telcos, auto lenders, and buy-now-pay-later platforms. Integration partners include Alloy, Persona, Provenir and Zoot, which is the polite way of saying SentiLink is wired into most of the places a fraud signal actually needs to land.
Three million verifications a day is the easier number to publish. The harder number, the one SentiLink does not put on a billboard, is how much fraud loss each of those calls prevents. Industry estimates for synthetic-identity loss alone hover near $20 billion annually. SentiLink's pitch is that they have measurably bent that curve for the lenders they work with - and several published case studies, dryly written, suggest they have.
§Make the financial system safe for real people.
SentiLink's stated mission is to make it safe and easy for consumers to interact with financial institutions. In private, the founders talk about it in narrower terms: every time the model catches a synthetic, a real loan dollar stays available for someone who actually needs it. The flip side is more interesting. Every time the model wrongly flags a thin-file immigrant or a teenager with a fresh SSN, a real customer is locked out. The team is unusually clear-eyed about that trade-off, which is rarer in the fraud industry than it should be.
§Deepfakes, AI agents, and the next decade.
The synthetic-identity problem is about to get harder. Generative models can now produce convincing identity documents in seconds. AI agents are about to start opening accounts on behalf of humans, which means the next fraud wave will not look like a person at all. SentiLink has been writing about this for two years - their annual fraud report has become required reading in risk circles - and their bet is that signal-based, behavior-based verification will age better than document-based verification.
If they are right, the company looks less like a startup and more like infrastructure. If they are wrong, $85 million was a very specific kind of mistake.
§Back to the quiet floor.
It is still a Tuesday on New Montgomery Street. Three million applications have moved through SentiLink's API since morning. A handful have been quietly stopped. The statisticians have not stood up to cheer. Nobody is going to tweet about it. The analyst who made the last manual-review call is already on the next one.
That is the company. A bet that financial trust is, mostly, a measurement problem. And that the right people to solve it are the ones least interested in saying so.
#fintech #fraud-detection #synthetic-identity #identity-verification #kyc #machine-learning #a16z #san-francisco #series-b