Building intelligent systems for evolving risk

Kris is working on tailoring model development strategy to meet the unique demands of our space. He was previously at JP Morgan Chase & Co where he led development of risk models in the Consumer & Community Banking space, and drove innovation initiatives involving secure, local language models. In his free time you can find him birding around the Bay Area, and volunteering for local habitat restoration efforts.

The successful business foothold landed by AI tools like ChatGPT and Claude has encouraged the assumption that, the best thing to do when faced with a big problem, is to reach for an even bigger foundation large language model trained to do more. But in financial technology, challenges like verifying identity and detecting fraud are not solved by asking one model for one score. These challenges require different kinds of holistic evidence. This includes purpose-built models, rules, and human judgment working together. That is the systems-first approach we take at Cardless. We build for the problem at hand, with an eye towards modular models and tools that can be repurposed across multiple business functions.

General-purpose chat models have changed what small teams can build, and we use them across our business daily as powerful force multipliers, but not as substitutes for subject matter expertise. At Cardless, our data science strategy and risk posture favor smaller, purpose-built models, letting us tailor our models and rules to the distinct needs and context of each brand partner. Today, these predictive tools are are even used in agent-driven workflows for investigations.

Our in-house fraud identity engine

Parallax, our in-house, fraud-identity engine, is one example of putting our philosophy in practice. The engine’s name is borrowed from the actual phenomenon of parallax, where an object appears to shift position between two different visual perspectives, for example holding up a finger in front of you and alternating closing your left and right eye while continuing to focus on it will give it the appearance of being displaced back and forth. In this same regard, a customer liveness check that appears benign initially may be revealed to have hallmarks of gen AI video, or injection into the KYC pipeline rather than it being naturally collected. The deepfake detection and identity resolution models in Parallax give us a layer of defense at application that also provides lift in catching coordinated risky spending behavior within an individual product or across the portfolio. Stronger identity resolution sharpens these fraud network signals. Investigations produce well-governed outcomes, and those outcomes strengthen the historical record available to the next generation of our models.

Seeing liveness in context

Diving into the first stage in Parallax, we apply the same “systems-first” view to KYC and liveness. A video is more than just a stack of images to classify, it carries facial geometry, scene context, motion, optics, device confidence indicators and traces of the capture pipeline too. In response, parts of our stack include checks on data tampering, Gen AI templating, and landmark tracking (is this person moving naturally? Or, they are moving naturally, it just happens to follow the same script as a burst of 20 other applicant videos that day?). New world-representation models such as DINO and V-JEPA help us answer the second question and go further still to get at the real identity question. Does this video make sense in context?  Rather than rely on any single decisive signal or model to answer, we look for collective agreement across views. Together, these perspectives create cross-domain evidence that is harder to spoof and easier for our analysts to interrogate.

Building the feedback loop

This work crosses computer vision, graph intelligence, data engineering, product engineering, and fraud strategy and analytics. The interesting challenge ahead will be building the infrastructure and feedback loops that let specialized models improve one another across the business, without sacrificing good governance or our agility in responding to new, AI-enhanced, or agentic, fraud events.

A new playbook for risk

At Cardless, that means creating purpose-built systems for the specific customers, card programs, and risk patterns we work with. If this type of open-ended problem solving excites you, we’re looking for others with cross-domain expertise who want to come design these greenfield, intelligent systems with us and contribute to a new playbook of risk strategy and fraud mitigation.