With limited impact on genuine customers and complaints
Pre-screened packages replace late fixes that could add a week or two
Fraud first, then credit, customer experience, and marketing
About Zempler Bank
Zempler Bank is a UK digital bank serving small and medium-sized enterprises (SMEs), sole traders, and consumers. It has held a full banking license since 2021 and built its reputation on secure, efficient banking for businesses that value simplicity and service. James Coveney, Head of Data Science, leads a team that develops algorithms for fraud detection, credit risk, and anti-money laundering (AML). Their hardest problem is spotting fraud with precision while keeping the customer experience smooth, as fraudsters grow more sophisticated and transaction volumes, which need real-time analysis, keep rising.
The team wanted cutting-edge machine learning, but its security and compliance requirements demanded strict governance over code dependencies in production. The tension was clear: move fast on innovation while keeping the standards banking regulators require. “With AI getting more complex, having loads of disparate data inputs and making sure that the algorithms we deploy are stable and perform as we expect is a challenge,” Coveney explains. “The only way to do that is to use data science. You have to use machine learning to deliver anything that is sufficiently complex.”
The Challenge: Outgrowing Legacy Statistical Software
Zempler’s team had relied on traditional statistical software for its critical modeling. It served well for credit scorecards and analytics, but became a limit once the team wanted more sophisticated machine learning. “We wanted to start using gradient boosted decision trees, which are very useful in our industry,” Coveney explains. “The support for those kinds of techniques in our previous platform was very limited.”
The proprietary software also tied the team to the vendor’s roadmap. “When you use an enterprise, closed-source language, you’re at the behest of the developers of that language,” he notes. “There will be a lot of times when you have to spend a large amount of time doing some bespoke solution, making the code do something that there’s not an inbuilt function for.”
Python’s open source ecosystem offered another path: new techniques as soon as they’re available. “With Python, it’s very unlikely that someone hasn’t solved a problem in a certain way,” Coveney explains. “You get all the benefits of open source: newer things are available very quickly, there’s a huge amount of innovation happening.”
The Stakes: Vulnerabilities Found Too Late
But Zempler used Python sparingly in production because of security concerns. The team needed the latest open source packages for more advanced fraud models, and its security requirements demanded strict control over package vulnerabilities. “The problem we had in the past was that our algorithms are, by necessity, developed on production data, so full access to open-source code was not an option within this environment,” Coveney explains. “Similarly, we did not want to leave security-scans to the implementation phase of the development.”
Scanning manually after development created bottlenecks and security gaps. Without proactive screening, developers often found vulnerabilities only after committing code, which meant dropping their work for days or even weeks of retrofitting. “It can be pretty material,” Coveney explains. “If somebody’s used a version of a package that has a vulnerability, and the fix is to upgrade to a version that has material changes, they might have to go and rewrite the entire codebase that they’ve written. It could add a good week or two onto something in the worst case scenario.” Even in simpler cases, the team faced “a fairly regular stream of half-day, one-day fixes,” constant interruptions that pulled analysts away from strategic work.
Why Anaconda: Pre-Screened Open Source
“What Anaconda gives us is the ability to pre-screen our development environment so that we know that when we release into production there are no vulnerabilities,” Coveney explains. Screening up front means vulnerabilities surface before release, which cuts the rework and the interruptions.
“It makes development much more secure for us, and because we have that confidence that what we’re going to release hasn’t got any vulnerabilities, it also removes human error,” Coveney notes. The team also saw efficiency gains in its daily workflows.
Anaconda also keeps the open source speed the team wanted. “Anaconda overlays the most important elements of enterprise security onto an open-source language, which means you retain access to cutting-edge tools while bringing in the security and governance elements that we need as a bank,” Coveney explains.
The Results: Fraud Down, Customers Undisturbed
With a secure development environment in place, the team could focus on sophisticated fraud models that protect the bank and its customers. Fraud detection means analyzing vast, disparate data in real time: transaction details, spending history, behavioral patterns, device information, and session data. The target is also rare, since less than 1% of transactions are fraudulent, and false positives disrupt legitimate customers. “You need to find that balance between fraud prevention and making the customer experience as smooth as possible, not declining too many legitimate transactions, and not having people sit at checkout ringing their bank to approve purchases,” Coveney explains.
Using machine learning models built on Anaconda, Zempler has reduced fraud by over 90%, with limited impact on genuine customers and complaints.
That protection matters most for small businesses, a core part of Zempler’s customer base. “Some of the most damaging fraud for small businesses is scams such as invoice redirect or CEO scams as well as account takeover,” Coveney notes. “Small businesses also sometimes have heightened risk of card fraud in instances where they may have multiple cardholders within the company, which increases the risk of the details being compromised.”
Speed matters too, because fraud tactics change quickly. “You do sometimes need to react quite quickly. You might want to implement a new algorithm or retrain an existing one and get it into production as quickly as you can,” Coveney notes. “That can only be done if you make the process as simple as possible.”
The Governance Advantage
Zempler’s CISO was part of the sign-off. “Our CISO was involved in the sign-off process,” Coveney recalls. “They were thrilled when I suggested it as a tool, because otherwise it would have been something we’d have to build manually. We would have had to basically do what Anaconda has done ourselves, which wouldn’t have been a valuable use of time.”
With Anaconda’s governance framework in place, Zempler gained the visibility and control to support its audit requirements while letting the data science team keep its speed.
What’s Next: AI Across the Bank
Fraud detection opened doors. Zempler now uses Anaconda across all its data science projects, with key initiatives in credit, customer experience, and marketing. “We have been able to use more advanced ML in areas such as credit, customer experience, and marketing where before we were limited in what we could automate,” Coveney explains. The roadmap includes further enhancement of underwriting and more machine learning across customer-facing and operational systems.
“The most important part of the process is how you are bringing AI/ML into production in a secure, efficient, and scalable manner,” Coveney reflects. “Anaconda has been very important in allowing us to ensure we have security embedded into our code from the onset of a project, making the transition to production much smoother.”
His advice to other data science leaders is to keep development and production close together: “I would recommend ensuring that the tools you use in development of ML/AI are as closely aligned to those you use to bring it into production as possible, in order to minimize the amount of re-coding involved in implementation. The smaller the gap between these two environments, the faster the implementation and the lower the chance of models being implemented incorrectly.”
For Zempler, Anaconda delivered the confidence to innovate with open source at enterprise scale, backed by the security and governance a bank requires.
Want to bring open source into production with the governance a bank requires? Contact our team for a personalized demo.
“They were thrilled when I suggested it as a tool, because otherwise it would have been something we’d have to build manually. We would have had to basically do what Anaconda has done ourselves, which wouldn’t have been a valuable use of time.”
James Coveney
Head of Data Science, Zempler Bank