CONFIDENTIAL CLIENT
How a technology company built an intelligent Pix fraud detection system with Machine Learning
Detecting Pix fraud in <1s with no labeled data for supervised training
Intelligent monitoring of Pix transactionsFraud detectionwith no reliance on labeled dataReal-time inference with millisecond response
In response to rising fraud on Pix, Brazil’s instant payment system, a technology company that specializes in banking software built, with BlueMetrics, a solution based on unsupervised machine learning to detect anomalies in real time. Without relying on labeled data, the system uses clustering techniques to learn the normal behavior of each account and flag suspicious transactions with precision and speed, processing each operation in milliseconds without affecting Pix response times.
Overview
The rise of instant payments in Brazil, Pix, brought unprecedented speed to consumers and businesses, quickly becoming one of the country’s main ways to move money. That shift, however, also opened the door to new kinds of fraud, increasingly sophisticated and hard to catch with traditional methods.
Facing this tough scenario, a banking software vendor decided to innovate and offer its clients, banks and fintechs, a new security layer based on artificial intelligence. The goal was clear: make sure fraud could be caught accurately and early, without hurting transaction response times, a critical factor in the Pix ecosystem.
The real challenge was balancing performance and intelligence: how do you detect fraud in real time without a labeled history of past cases, which is the common situation at financial institutions? The answer called for an innovative approach, one that could learn from account behavior and react quickly to patterns that fall outside the norm.
Market context:
- Rapid growth of Pix and banking digitization
- More fraud attempts happening in real time
- High demands on transaction performance and security
Problem: how do you get millisecond precision with no labeled data?
Pix sets a maximum of 40 seconds for a transaction to complete. That means any anti-fraud analysis has to be extremely fast, efficient, and, above all, integrated transparently into the operation. Making the challenge harder, the company had no labeled dataset with examples of fraud, a common situation in banking, where fraud often is not documented in the detail supervised model training requires.
On top of that, each bank account has its own behavior patterns, which vary by client type (individual or business), transaction profile, days and times of activity, and other factors. In that context, fixed rules simply could not capture every nuance and exception, and could even create false positives or let suspicious transactions through.
The situation called for an intelligent approach that could learn from the data and keep adapting to different usage profiles. “This is exactly the kind of challenge that motivates us here at BlueMetrics: it’s strategic for the client and has the potential to deliver concrete, measurable results even in the short term. With a well-designed solution, you can combine intelligence and speed without giving up reliability.”
Main challenges:
- Operational limits:
- Business limits:
- Technology limits:
The solution: anomaly detection with unsupervised Machine Learning
With support from BlueMetrics, the company deployed an unsupervised machine learning model built specifically to identify behavioral anomalies in high-volume transaction environments. The lack of labeled data called for a clustering-based approach, in which the system learns, on its own, the typical movement patterns of each account, considering variables such as transaction frequency, value, and time, and then compares each new operation against that history, measuring its statistical “distance” from expected behavior.
This behavior-driven architecture was essential to catch subtle deviations that could still point to fraud, without relying on fixed rules or preset exception lists.
The solution was built on native AWS technologies, ensuring scalability, security, and high availability, and it added real-time inference mechanisms that classify transactions in milliseconds. Each transaction is analyzed automatically and gets a percentage risk score, enabling immediate decisions. All of this without compromising the response time the Central Bank requires for Pix settlement.
That speed, combined with the model’s statistical precision, gave the company’s bank and fintech clients a meaningful competitive edge: a security layer that is effective, discreet, and fully integrated into the user’s journey.
Main components:
- Unsupervised machine learning model
- Clustering techniques for per-account behavior analysis
- Real-time inference with Amazon SageMaker
Technology differentiators:
- A solution built on 100% cloud architecture (AWS)
- Anomaly detection with no need for labeled data
- Response timeunder 1 second
Immediate benefits:
- Fraud identifiedbefore the transaction completes
- Continuous adaptation to new behavior patterns
- Lower financial losses and stronger confidence in the system
Results:
The system began flagging suspicious transactions in under 1 second, which allowed real-time alerts, before an operation was even settled, and gave partner institutions the chance to act preventively. That ultra-fast response was decisive in protecting users and keeping the system sound, especially amid Pix’s exponential growth.
In simulations with historical data, the feature was estimated to have avoided up to R$ 1.5 million in losses, proving its potential for direct impact on clients’ financial results.
But the benefits go beyond loss mitigation. The solution added strategic value to the portfolio of the banking software developer, which now offers not just a management tool but an intelligent, proactive security infrastructure.
The new capability raised the perceived value of the platform, increasing its competitiveness and reinforcing the brand’s position as a reference in anti-fraud innovation and technology in the Pix world. The mix of technical performance and concrete results made the feature a genuine competitive edge.
he adds: “It’s very rewarding when we deliver solutions that create real, immediate value for the client, solving concrete problems with direct impact on results. That’s exactly the kind of challenge that drives us.”
Impact on operations:
- Automated identification of suspicious transactions in milliseconds
- Fewer financial losses thanks to real-time fraud alerts
- A stronger value proposition for the banking software with a new security layer
Technology advances and integration:
- An unsupervised model adapted to different client profiles
- Transaction processing with real-time inference
- Transparent integration with banking infrastructure without compromising Pix timing
Technologies used
The solution was designed using several AWS technologies, including:
AWS services
- Sagemaker
- S3
Languages, libs, and frameworks
- Python
Conclusion:
In this case, artificial intelligence was not just an ally: it was the real engine of innovation. By combining unsupervised machine learning with a robust cloud architecture, the company built a solution that meets the most critical requirements of the financial sector: precision, speed, and scalability. With the ability to flag suspicious transactions in milliseconds and a potential prevention rate that could have avoided up to R$ 1.5 million in fraud, the feature goes well beyond an extra security layer.
It significantly improves the user experience, protects millions in assets, and strengthens the core product of the developer, which now stands out in the market for offering an intelligent, real-time anti-fraud solution.
More than solving a technical problem, applying AI here turned an operational bottleneck into a strategic asset, and that is what makes this kind of innovation so valuable: its ability to turn complexity into a concrete competitive advantage.