This Startup Catches Fraud at Scale

Key Concepts

• AI Agents: Autonomous software entities capable of reasoning, using tools, and executing complex workflows (e.g., fraud detection, identity verification) without human intervention.
• Stealth Mode: A business strategy where a company operates without public disclosure to protect sensitive intellectual property or competitive advantages.
• KYC/KYB (Know Your Customer / Know Your Business): Compliance processes used to verify the identity of individuals and the legitimacy of business entities to prevent fraud and money laundering.
• Deterministic Systems: Traditional rule-based software (e.g., "if transaction > $1,000, flag it") that lacks the flexibility of AI-driven reasoning.
• Self-Healing Systems: Architectures that can adapt to new fraud patterns and evolve without requiring manual updates to rules or classifiers.
• Unstructured Data: Information that lacks a pre-defined data model (e.g., images, bios, web content), which Variance’s agents are designed to process and analyze.

1. Company Overview and Funding

Variance, co-founded by Karine Mulada and Michael, has officially emerged from three years of stealth mode, announcing a $21 million Series A funding round. The company provides purpose-built AI agents for risk and compliance, automating high-stakes tasks such as content review, fraud detection, and identity verification for Fortune 500 companies and major marketplaces.

2. Real-World Applications and Case Studies

• GoFundMe: Variance automates the verification of fundraisers to prevent fraud. For example, during crises, the system distinguishes between legitimate family-led fundraisers and fraudulent actors attempting to capitalize on public sympathy.
• Marketplaces/Gig Economy: The platform verifies identities for delivery drivers and service providers by analyzing selfies and government-issued IDs against company standard operating procedures (SOPs).
• Complex KYB: Variance maps complex business relationships, identifying shell companies and potential links to sanctioned countries or adverse media, a task previously requiring massive manual effort.
• Content Moderation: During sensitive periods like elections, Variance’s agents detect sophisticated, state-sponsored fraud rings by analyzing entity relationships across vast datasets, preventing the spread of misinformation and potential physical threats.

3. Methodology: The Three Building Blocks

Variance utilizes a streamlined framework for building AI agents:

1. Compliance Documents/SOPs: The foundational rules and requirements provided by the client.
2. Tooling: Custom-built interfaces that allow agents to interact with internal data stores and external sources.
3. Data Access: The ability to ingest both internal unstructured data and external open-web data to provide context for decision-making.

4. Technical Challenges and Innovation

• Data Ingestion: A primary challenge was accessing data scattered across 5–10 different systems, often hidden behind legacy UIs. Variance developed agents that can "spin up a browser" to scrape data from legacy tools, effectively mimicking a human analyst’s workflow.
• Moving Beyond Deterministic Systems: Unlike traditional fraud systems that rely on rigid rules and isolated classifiers, Variance uses LLM-based agents that can reason over unstructured data, allowing for a "self-healing" system that adapts to new adversarial tactics in real-time.
• AI Coding Maximalism: With a lean team of five engineers, Variance operates as an "AI-first" company. Every engineer manages a small team of coding agents, effectively multiplying their output to that of a 25-person team.

5. Key Arguments and Perspectives

• The "Shadow" Strategy: Karine argues that because their work involves sensitive security data, maintaining a low profile is a strategic necessity. Publicizing their methods could inadvertently provide a roadmap for bad actors to bypass their defenses.
• The "Founder-Led" Resilience: The founders emphasize a "sense of duty" to solve the specific problem of fraud, which they believe is the primary reason they persisted through challenges, including a severe accident that left the CEO hospitalized.
• Human-in-the-Loop: While the goal is automation, Variance acknowledges that the final 1% of complex cases requires human intervention. Consequently, they have invested heavily in building high-quality investigative dashboards for human analysts.

6. Notable Quotes

• "We're building the systems that are often used by the bad guys, but we're building them for the good guys." — Karine Mulada
• "Every engineer is going to have three monitors with their coding agents running... in terms of output, everyone is a manager of a small team of AI agents." — Karine Mulada

7. Synthesis and Conclusion

Variance represents a shift from static, rule-based compliance to dynamic, agentic AI systems. By focusing on the "burning" problem of fraud and compliance, the company has successfully transitioned from a niche startup to a critical infrastructure provider for global enterprises. Their success is rooted in a deep understanding of the fraud landscape, a commitment to "self-healing" architectures, and an aggressive adoption of AI-driven development workflows. The company’s future focus remains on scaling their operations while maintaining the high-precision decision-making required for enterprise-grade risk management.