This Startup Catches Fraud at Scale

Predictions and Expectations:

• Karine expects that beyond the Series A, Variance will "always be a company that's a little bit more in the shadows."
• She believes that without AI agents, tracing abuse graphs across unstructured data on the open web is impossible for human analysts.
• She anticipates that the current patchwork of rules engines, classifiers, and humans will be replaced by a "fully self-healing system" where AI agents handle the entire workflow without needing separate classifiers or human reasoning for most cases.
• She expects that with coding agents, the current five-engineer team will effectively produce software output closer to a 25-person team.
• Karine expects that the need for human analysts to manually Google names and apply judgment will be eliminated as agents can now directly query data stores and reason over unstructured web data.
• She anticipates that AI coding agents will enable non-technical customer success managers to directly ship features to enterprise customers within a few hours without engineering intervention.
• She expects the system to evolve rapidly, allowing companies to ship faster and open new product lines without the bottlenecks of traditional human-in-the-loop fraud systems.
• Karine expects to detect and prevent physical violence by identifying online threats made with plans for physical harm at scale.
• She believes that the ability to detect complex fraud rings involving state-sponsored actors during elections would not have been possible with isolated classifiers.
• Karine expects that the "bus factor" risk associated with founder-led sales and operations will necessitate scaling the founding team to ensure resilience.
• She is not sure how the cost structure and performance of their models will evolve, noting that OpenAI's release of new models during their pilot changed their cost structure by a "10x factor."
• She expects that the industry will eventually require a more resilient and efficient way to solve fraud and compliance problems, similar to the approach taken at Apple.

Timelines and Milestones:

• Variance has been building in stealth for "the past three years."
• The company recently announced a "$21 million Series A."
• In "July 2024," the company experienced a period where revenue was "doubling within the month" and then "doubling the month after."
• Karine states she was hospitalized for "about 10 days" after being hit by a truck.
• She was unable to walk for "about ten days" following the accident.
• She was "out of commission" for about a week, with the company facing potential disruption for "a year and a half" as the CEO recovered.
• It took "eight months" to land their first customer, IAC.
• GPT-4 was released "during the batch" when the company was running its first pilot.

Technology and Product Direction:

• Variance plans to hire across the board, specifically for "back-end" and "front-end" roles.
• The company is shifting from a pure API call model to building a "full end-to-end decisioning layer" that includes a "really good dashboard" and "investigative visual tool" for the remaining 1% of complex cases requiring human review.
• Variance is introducing a third integration method: spinning up a browser to "scrape from a UI that was built for a human" to reason over data hidden behind interfaces.
• The company plans to continue leveraging "coding agents" to maximize the output of its small engineering team.
• Variance aims to maintain a "fully automated" decision-making process for content, fraud, and identity reviews at scale.
• The company intends to continue using AI agents to verify identity, content, and complex KYB (Know Your Business) verifications, including graph skill construction for shell companies.
• They plan to utilize access to "over a hundred business registries across the world" and the "open web" to support international compliance investigations.
• The team plans to pool "unstructured data" into their own data stores to enable AI agents to reason over it.

Market and Industry Outlook:

• Karine expects that customers in the trust and safety space, such as marketplaces and gig economy platforms, face a constant "cat and mouse game with the bad guys" where they need "secret weapons" that cannot be marketed openly.
• She anticipates that the compliance market will continue to demand scalable solutions for complex identity reviews and KYB verifications that are currently performed "entirely manual."
• She expects that the "dynamic environment" of fraud, driven by adversarial actors, requires a system that can evolve "really rapidly."
• The company believes that the industry is moving away from deterministic rules and classifiers toward "agentic systems" that can materialize features on the fly.
• Karine expects that the demand for content review and fraud detection will remain high, particularly in "crisis driven" scenarios like natural disasters or elections where fraud spikes.
• She anticipates that customers like IAC and GoFundMe will rely on Variance to handle compliance risks that are "semi-impossible" to map to traditional classifiers or regular expressions.

Company Plans:

• Variance plans to scale the founding team ("scale me") to prevent future disruptions if the CEO is unable to work.
• The company is hiring to support its growth from 12 employees (including five software engineers) to a larger operational capacity.
• They plan to continue operating with a "very, very lean" team structure while maintaining a strong "ownership culture."
• The team intends to remain "in person in San Francisco every day."
• They plan to continue selling to "fortune 500s" and large marketplaces, maintaining their focus on enterprise clients.
• Variance plans to keep the majority of its operations "in the shadows" due to the sensitive nature of the fraud and abuse patterns they detect.

Financial Guidance:

• Variance has raised "$21 million" in Series A funding.
• The company has previously experienced a "step function" in growth where revenue doubled "within the month" and doubled again "the month after" in July 2024.
• The company has been operating with a "very, very lean" cost structure, initially consisting of only five software engineers processing "petabytes of data."

Risks and Caveats:

• Karine warns that if Variance were to market its specific use cases, it "may create more fraud" or "abuse" because adversaries would learn from the disclosed defenses.
• There is a risk that the company relies too heavily on the founders ("bus factor of one"), which was highlighted when Karine was injured and Michael was left managing sales and customer relationships alone.
• The company faced a moment of existential risk when Karine was hospitalized, with Michael worrying that the company might "part ways" or end entirely.
• There is a risk that the dynamic nature of fraud environments could outpace the system if it does not have a tight feedback loop.
• The company notes that data is often scattered across "five to ten different systems" and hidden behind UIs, creating significant technical challenges in data aggregation.
• Karine mentions the risk of "misinformation" and "abuse vectors" having "serious physical implications," noting that detected threats can involve plans for physical harm.
• The company acknowledges that building the system is "really hard" and that access to the open web was one of the "final nodes that made this whole problem really hard to automate."

Confidence and Disagreement:

• Karine expresses high confidence that the "AI agents are able to close the loop from a reliance and self-healing standpoint," removing the need for humans in the loop for 99% of cases.
• She is "really really really stubborn" about ensuring the system has "no nodes that were inefficient."
• She confidently states that the company will "always" remain somewhat in the shadows.
• Karine expresses strong conviction that her specific skills in fraud and the industry's current problems provided a "strong sense of duty" to build this product.
• She is confident that the shift to agentic systems allows for "self-healing" systems that can "thrive in a dynamic environment."
• She states with confidence that "no human team of fraud analysts could ever have been able" to detect the sophisticated fraud rings during elections that the AI agents identified.
• Karine expects that the team's culture of high ownership and agency will continue to drive success.
• She is confident that the "coding agents" will effectively make every engineer a "manager of a small team of ai agents."