Dhiraj Bhat

Product manager. Builder. Fraud nerd.

Seven years building products at the intersection of financial crime and AI. I've helped 100+ financial institutions share fraud intelligence without sharing a single piece of customer PII, shipped AI into audited investigation workflows, and argued in front of FinCEN about what "too many alerts" actually means. (The answer depends on whether you're in fraud or AML. The full answer took twenty minutes.)

I also co-invented a patent for privacy-preserving fraud network architecture, vibe-coded a fraud simulation browser that unlocked ~$150K in partner ARR, and built a job search coach to coach myself through interviews. It worked. Alas, not novel.

Currently: Lead PM at Interface AI. Previously: Senior PM / GM at Unit21.

Career

Scope & Responsibilities
Interface AI Unit21 Emergent / Qenta Duke Bosch Rexroth
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026

Selected Work

The Right Product Battle to Lose

In 2022 I argued to save Unit21's identity verification product. It was bleeding support tickets from upstream vendor outages, sales was avoiding it for margin reasons, and the engineering lead wanted a full backend rebuild. I had the retention data, a PLG motion sketched out, and a strategic argument: you can't be downstream in fraud without owning the onboarding journey. I lost. Leadership killed IDV on the economics.

About a year later, Unit21 launched sanctions screening. Customers immediately asked: "Why don't you also do verification? Those two go together." The market validated the thesis I had documented. The company had already given up the position.

Startups optimize for margin when they should optimize for surface area. Losing well means documenting your reasoning so when the market comes back around, you can move fast instead of re-litigating the debate.

9 to 100: The Fraud Network Cold Start

Two years into building the Unit21 Fraud Consortium, usage hit a wall. Members spiked at onboarding then went quiet. The contribution model required confirmed fraud labels, meaning members had to complete a full investigation before feeding any signal back. For AML-heavy teams whose workflow ends at filing a SAR, that was asking them to do more work for a network not yet generating enough return to justify it. Classic cold-start trap.

I made two counterintuitive calls: accept True Positive alerts instead of confirmed fraud (a deliberate precision-for-recall tradeoff at the contribution layer), and make the data configurable rather than delivering a verdict. True positive alert counts went from 9 to over 100 in 2.5 months.

In a shared network, adoption is the product at early stages. Signal quality is a second-order problem. If you demand high-quality contributions from customers who aren't operationally ready to provide them, you kill the network before it ever gets to quality.

Shipping AI Into Regulated Fraud Workflows

The surface concern when Unit21 deployed AI agents to fraud investigation was "it'll take our jobs." The real concerns were auditability, legal exposure, and explaining AI decisions to regulators. I solved it in two places most PMs wouldn't touch: at the architecture layer (chose AWS Bedrock over direct API calls to eliminate the Data Processing Agreement requirement that would have killed the pilot), and at the agreement layer (reworked the consortium contract structure with our GC, from 60 pages to 4 with equivalent protection).

65 customers. 35-65% handle time reduction. Zero exits for legal, audit, or product reasons.

Legal friction is a product design problem, not a legal department problem. The fix is almost always upstream of where most PMs would touch it.

Why Banks Can't Leave Bad Software

Unit21 was losing deals to Verafin, the dominant fraud/AML incumbent for FIs. The sales team's theory: Verafin won on product features. Through direct discovery with FI compliance teams, I learned the real dynamics. Verafin won because auditors trusted it and because core banking providers (Jack Henry, FIS, Fiserv) actually run the relationship. Banks are terrified of their own core. Asking for file access triggers a ticket queue measured in months.

Then NASDAQ acquired Verafin and raised pricing 2-3x. That was the crack. I used it to reshape Unit21's FI go-to-market: stop trying to route through core banking, route around it through digital banking platforms instead. One integration, dozens of FIs. That became Unit21's primary FI motion.

The vendor you think you're competing with isn't winning on product. They're winning on audit friction and twenty-year incumbency. You don't beat that head-on. You route around it.

Writing

Recognition