Custom crypto-native machine learning models

Many fraud platforms train on consortium data and return a score with little visibility into what drove it. Oscilar instead trained Uphold’s models on its own data and customer journey, with Uphold defining the signals and outcomes that mattered. “A lot of fraud platforms were designed from an e-commerce perspective,” McCall said.

“We needed something crypto-native, tailored to our customer journey: onboarding, authentication, the transaction, and then how customers hold funds and keep transacting. All of that is unique to Uphold.”

Rob McCall

VP of Fraud Prevention

Uphold

That meant distinguishing stolen-payment fraud from first-party fraud and scam victims, and incorporating signals Uphold’s analysts already trusted. Device and behavioral intelligence added another layer. “Oscilar let us take all these signals we used to have in silos: IP, device, keystrokes, how things are input, and combine them,” McCall said. “You’re not looking at any one signal. You’re looking at the whole picture, and that’s what the model does.”

Uphold also controls how those models act on that intelligence. McCall’s team can tune thresholds to each product’s risk appetite, balancing fraud capture, review volume, and approval rates without waiting on engineering. “What stood out with Oscilar was the explainability, and the customization and agility we could have in developing all of our flows,” McCall said.

The clearest proof came in the early stages of a new fraud trend, when suspicious activity began appearing across accounts that looked unrelated on paper. Oscilar’s device and behavioral team connected the activity through harder-to-mask signals, then worked alongside Uphold’s fraud analysts to stop it before it spread further.

This was not a one-off. Uphold and Oscilar have continued to work this way as new fraud trends emerge, often with Oscilar reaching out proactively when its team spots a pattern worth investigating. As fraud fragments across more accounts, devices, and identities, that combination of connected signals and hands-on investigation helps Uphold surface patterns that individual accounts can hide.

How the fraud function became a growth enabler

The broader results showed up quickly. Average handle time for fraud alert review dropped 30% in the first three months, across the full review from alert to decision. False positives fell 15%, and overall fraud losses declined 35% year over year.

More important was what that performance changed for the business. Fraud had been a potential constraint on new products; with Oscilar, McCall says that calculus flipped.

“[Oscilar] allows Uphold to think about offering any kind of product without worrying whether we have the right controls in place. We can sit in the background and never be a blocker to growth. We can actually amplify it.”

Rob McCall

VP of Fraud Prevention

Uphold

Better detection also gave Uphold room to approve more good customers. “The revenue we’ve captured has far outweighed the return losses,” McCall said. And because the decisions are explainable, the team can defend that wider funnel. “We can explain the why on anything we do, whether it’s the friction we’re putting in, the controls, or the model’s decision.”

The enterprise API makes that impact concrete. Each new client brings a different customer base and flow; on a rigid platform, that could mean bespoke controls or a one-size-fits-all gate that slows the sale. On Oscilar, fraud controls become part of onboarding. “Oscilar lets us adapt and customize fraud controls for every enterprise client,” McCall said. “We’re not locked into a box.” Fraud stopped being the reason a new revenue line had to wait.

The same flexibility matters when crypto volume spikes without warning. A twenty percent market move can send trading, onboarding, and review queues surging in an afternoon.

“A market rush used to mean operations pain. Now we take the volume instantly, and the customer never feels it.”

Rob McCall

VP of Fraud Prevention

Uphold

More than 150 people at Uphold now work in Oscilar across fraud and compliance. Some 30 live workflows are managed day to day by Uphold analysts. Policy and operations teams can build, back-test, and deploy strategies themselves, and when a new typology appears, the response can be a workflow change the same day.

One platform closes the fraud-compliance gap

Walton watched the fraud deployment before committing his own program.

“The power of Oscilar’s automation and AI was clear early on. It wasn’t bolted on; it’s a core competency, and you can see that in the product.”

Frank Walton

Global Chief Compliance Officer

Uphold

For compliance, the opportunity was bigger than automating alerts. Volumes of alerts had been worked by people across multiple levels of review, and every case that moved from fraud into compliance crossed a system boundary with them.

Oscilar closes that seam. “An alert generates in one team, gets escalated, and comes to compliance, and it all happens inside Oscilar,” Walton said. “An investigation that used to pass through three or four sets of hands from alert to SAR filing can be handled by a fraud person and a compliance person, with the context carrying across.”

That continuity removes the errors that creep into handoffs. “Something as simple as the numbers being right in a SAR filing. That shows up on almost every institution’s audit findings, and it’s preventable when everything lives in one system,” Walton said.

That is FRAML in practice: the same customer, evidence, and history carried from fraud review into compliance action instead of being rebuilt on the other side. Fraud and compliance become parts of one decision chain, with analytics following the same path.

The shared view also helps Walton anticipate what is coming. “If Rob’s team is seeing an influx in fraud alerts, we can reasonably expect an influx in escalations. That helps us prepare.” And because the platform absorbs that volume, the business no longer has to plan compliance capacity around every possible spike. “With Oscilar, because it scales, we don’t need those knee-jerk reactions,” Walton said. “The business can assess a product’s P&L knowing the compliance cost won’t materially change.”

For Walton, that is where the operating model is headed. “Oscilar will be a key component to making Uphold’s Compliance team a more scalable and efficient department.” The regulatory expectations are moving in the same direction. “When I started my career, everything was about how many people you had reviewing everything, what their CVs looked like, whether they knew what they were doing,” he said. “Now there’s almost more skepticism about doing it manually than not. What regulators want to see is that you’re being strategic. Do you have AI and machine learning working in the background so your people are looking at the things that really matter?”

With fraud and compliance finally operating from the same foundation, specialized agents could begin working across both.

Uphold made an early bet on Oscilar’s agents

“The agents were the reason we chose Oscilar,” McCall said. “Eighteen months ago nobody else was building them, not for alerts, not for workflows, not for fraud trends. We bought the vision, and they’ve delivered on it.”

Oscilar had built the foundation first: Uphold’s customer data and decision history, custom models, device and behavioral intelligence, workflows, and cases in one system. That gives each agent more than the alert in front of it and lets specialized agents work from the same institutional context. “Automation and scalability are the two words I’ve been pushing. Oscilar gives us both, and the agents are how we get there,” McCall added.

On the fraud side, an L1 review agent works the first pass of an alert using Uphold’s procedures and broader account context, while a case narrative agent drafts the investigation write-up. “It’s exactly how our analysts work an alert,” McCall said. “We focus on the event that triggered it, but we consider the customer’s whole account. The agent does the same thing, from the same context, faster.” Early results showed alert resolution time falling roughly 35%.

“The agents were the reason we chose Oscilar.”

Rob McCall

VP of Fraud Prevention

Uphold

On the compliance side, sanctions and PEP screening agents apply Uphold’s own dispositioning procedures. “It was imperative that the agent matched our procedures, not the other way around,” Walton said. “I would not have been comfortable with an out-of-the-box agent.” He set a conservative bar: “We only wanted it closing things we were totally certain were safe to close. If it had missed an OFAC match, that would have been a dealbreaker.” The sanctions agent flagged every true match with zero critical misses. Document extraction is already in use in due diligence, with further due diligence and KYC capabilities following. Up to 5x faster screening and EDD document review is projected.

The bigger advantage is orchestration. Specialized agents can draw on the same customer history, signals, policies, cases, and prior decisions, while analyst feedback adds to the institutional intelligence available across the operation and humans remain in the loop.

“We’ve done 100% QC on every alert an agent has worked,” Walton said. “It’s reliable and it’s provable. The next step is segmenting the alerts and giving the agent more discretion where it has earned it.” That discretion starts with repetitive, high-volume work.

Walton knows that work firsthand. “An agent doesn’t get bored,” he said. “I’ve done level one reviews myself. You do it for a year and you think, there has to be something better than this.” McCall sees the same shift at the operating level. “With Oscilar, our front line review and immediate assessments will be automated, and we’ll focus on the true outliers, protecting the business with minimal impact to the good customer.”

Over a year after the initial bet, McCall says it has paid off. “Oscilar has delivered on everything it set out to build, and the agents are the latest proof. My confidence has only grown.” And part of that confidence came from how the two teams built together.