This Startup Secretly Detects Fraud For Fortune 500s

Variance is announcing its emergence from stealth alongside a $21 million Series A, marking the first time the company has publicly discussed its work powering trust and safety systems for some of the world's largest platforms. The company builds purpose-built AI agents that automate content review, fraud detection, and identity verification at scale for Fortune 500 companies, major marketplaces, and platforms processing millions of transactions daily.

The secrecy wasn't a marketing ploy—it was operational necessity. Variance deals with highly sensitive abuse vectors, and publicizing detection methods would create a roadmap for fraudsters. As Karine puts it, «We're building the systems that are often used by the bad guys, but we're building them for the good guys.» Customers don't want adversaries to know what their defenses look like, so Variance has operated as the invisible infrastructure layer behind products millions of people use every day.

Despite the stealth posture, the company has quietly scaled to serve household-name platforms. They review every GoFundMe fundraiser, verify sellers for a Fortune 50 marketplace, and detect coordinated misinformation campaigns for politically exposed communities—all with a team of just twelve people, five of whom are engineers.

Variance has five software engineers processing petabytes of data and making automated decisions for some of the largest companies in the world. The team achieves this leverage by treating AI coding agents as force multipliers: every engineer runs three monitors with coding agents, effectively managing a small team of AI contributors. The company estimates their actual software output rivals that of a 25-person engineering org.

The culture is high-ownership and deeply collaborative. Karine and Michael don't dictate features—they give engineers problems, and those engineers own the solution end-to-end. Luke, an early engineer from Meter, now understands LLM evaluation better than most people in the industry. The team has two founders among twelve employees, and everyone operates with founder-level agency.

Even non-technical roles are augmented by AI. The customer success manager, who has no engineering background, now takes simple feature requests, feeds them to Cursor, and ships updates autonomously—then tells enterprise customers hours later that the feature is live, without ever looping in the engineering team.

Karine and Michael met as engineers on Apple's fraud team—Karine in data engineering, Michael in machine learning. They had a symbiotic working relationship: Michael's ML models were deployed via Karine's streaming jobs, serving fraud detection across iMessage, iCloud, and other Apple services. They saw firsthand how fraud systems relied on a patchwork of deterministic rules, narrow classifiers, and slow human review—a feedback loop too sluggish to keep pace with sophisticated adversaries.

They believed the problem could be solved in a more resilient, self-healing way, but the right vehicle was a company, not an internal project. Karine suggested they apply to Y Combinator. The motivation wasn't to start «a company»—it was to solve this specific problem with their rare, complementary skill sets. That sense of duty, and the conviction that large language models could finally close the automation gap, became the foundation of Variance.

GPT-4 launched during their YC batch, transforming cost and performance mid-pilot with their first customer. They were building in a world that was radically dynamic—and they were exactly the right team, at exactly the right time, to ride that wave.