Financial Technology · AI & Automation
Building a Real-Time Fraud Detection and Investigation Platform
An end-to-end AI-powered platform that reduced fraud losses by 38% while cutting analyst review time from 6 minutes to 90 seconds—without slowing down a single legitimate transaction.
Reduction in fraud losses
38%
Timeline
5 months
Industry
fintech
Type
Full-stack platform build
✦ About Our Partner
Our partner in this project is Digital Banking Platform, a fintech company building at scale.
Company Overview
Consumer fintech, venture-backed, 400K+ active accounts
Project Timeline
5 months
Engagement Type
Full-stack platform build
Discover the challenge and how we solved it
Continue →01
The Challenge of Scale
A rapidly scaling fintech platform faced an existential problem: their fraud controls couldn't keep up with growth.
Digital Banking Platform had grown to serve 400K+ consumer accounts with 2M+ monthly transactions. Their static fraud-prevention system—built when they were processing a fraction of current volume—was failing on two fronts: analysts spent 6+ minutes per flagged case with no visibility into why transactions were blocked, and the rule engine's reaction time meant legitimate customers were getting declined at checkout.
The operational toll was real. A lean engineering team spent days fighting fraud spikes. Every threshold adjustment created new problems. By the time Loam was engaged, incremental fixes were no longer enough.
Key Metrics
400K+
Active Accounts
2M+
Monthly Transactions
$180K
Monthly Fraud Loss
6 min
Avg Review Time
02 — Challenge & Goals
Static rules are inherently reactive. By the time you adjust a rule, fraudsters have already moved on. At scale, reactivity is a liability.
Understanding both the business and technical landscape
Business & Operations
The existing system evaluated transactions in isolation using hard-coded rules. Fraudsters operate across multiple accounts and sessions. The rules engine had zero visibility into these patterns. Each rule update required a full engineering cycle, meaning the team was always reacting to fraud that had already happened.
Technical Constraints
Fraud analysis happened on the critical checkout path. A slow fraud check meant slow checkout. At 2M+ monthly transactions, even a 100ms delay translated to abandoned carts and lost revenue.
Success Goals
Sub-200ms Decisions
Score every transaction in under 200ms without impacting checkout
Explainable AI
Show analysts exactly which signals drove each fraud decision
Configurable Control
Update thresholds in minutes, not weeks
Compliance-First
Maintain PCI-DSS integrity across all data flows
04 — Discovery & Strategy
Engineering Discovery
Before writing code, we spent two weeks analyzing 18 months of transaction data, shadowing analysts, and mapping system constraints. What we found shaped the entire platform.
Workflow Audit
Finding
Analysts were getting only a rule ID and amount. They had no visibility into which signals had fired, forcing 20+ minute manual investigations per case.
Decision
We chose gradient-boosted models over deep learning specifically for explainability. A deep-learning model would be more accurate, but unexplainable.
Outcome
With feature-attribution explanations, average case resolution time dropped from 6 minutes to 90 seconds.
Stakeholder Interviews
Finding
Real fraud clustered at the device and session level. A single attacker using the same device would attempt multiple transactions within minutes. Existing rules evaluating transactions in isolation missed this entirely.
Decision
Build a feature store that pre-computes device fingerprints, session behavior, and clustering. Make these signals available via fast lookup instead of expensive runtime computation.
Outcome
Pre-computed features became the single biggest leverage point for hitting the 200ms latency target.
Data and Risk Mapping
Finding
Checkout was synchronous. Any delay in fraud analysis translated directly to checkout latency. At traffic peaks, a single-threaded analysis would bottleneck.
Decision
Design the scoring system asynchronously. Fraud check runs in parallel with the transaction; decisions are used for post-transaction enforcement and alerts, not real-time blocking.
Outcome
Checkout latency was decoupled from fraud-scoring latency. Checkout stays fast regardless of model performance.
Product and Architecture Workshop
Finding
Identity signals (KYC provider data) and behavioral signals (device, session, transaction) lived in separate systems. Neither had a unified view of customer risk.
Decision
Build a feature store that unifies signals from identity, KYC, and behavioral systems. Carefully scope to ensure cardholder data never mixes with identity provider data inappropriately.
Outcome
The model gained visibility into the full customer risk picture without violating PCI compliance boundaries.
Validated Delivery Roadmap
Finding
The team needed a clear path from research to production with multiple validation gates to prevent shipping a model that looked good in testing but failed in real traffic.
Decision
Phase the rollout: shadow mode validation → gradual traffic ramp → real-time monitoring. Build automated drift detection to catch model degradation.
Outcome
Three weeks in shadow mode caught model drift before production. Gradual rollout identified edge cases in real traffic.
05 - System Architecture
System Architecture
A layered system where each layer serves a specific purpose. Responsibilities are separated so each can be optimized independently.
Experience Layer
Analyst review console and customer-facing adaptive authentication. Decoupled from core transaction systems so fraud-team deploys never risk checkout.
API & Orchestration
Dedicated fraud-scoring API called asynchronously. Handles model inference, feature lookup, and decision logging without touching the checkout path.
Rules & Intelligence
Pre-computed feature store with device fingerprints, session patterns, and behavioral clustering. Real-time path does fast lookups instead of computing features at request time.
Data & Integrations
Receives signals from identity verification, KYC, and transaction systems. Normalizes and pipes into the feature store with careful PCI scoping.
Security & Observability
Auto-scaling infrastructure with real-time monitoring of model accuracy, latency, and compliance metrics. Automated drift detection and alerting.
System Design Philosophy
Each layer serves a specific purpose and can be scaled or updated independently. The separation of concerns ensures that fraud scoring (layers 2-4) can be as complex and slow as needed without affecting the checkout experience. Checkout always stays fast because it's completely decoupled from the analysis pipeline.
06 — What We Built
Platform Architecture
The platform comprises four major systems that work together to detect fraud at scale while keeping checkout fast and analyst review time low.
Each system is designed to operate independently while contributing to a unified fraud detection and prevention strategy.
Real-Time Decision Engine
System 1
Unified Investigation Workspace
System 2
Rules & Workflow Control
System 3
Analytics, Audit & Reporting
System 4
Real-Time Decision Engine
Every transaction gets a risk score in under 200ms without touching checkout. Scores are explainable and actionable for analysts.
- ✓Sub-200ms latency via pre-computed features
- ✓No impact to checkout experience
- ✓Full feature-attribution visibility
- ✓Version-controlled model deployment
Unified Investigation Workspace
Analysts see the complete picture. Every flagged transaction shows the exact features that drove the fraud score, plus customer history and risk context.
- ✓Feature-attribution explanations (SHAP values)
- ✓One-click threshold adjustments
- ✓Real-time case queue
- ✓Color-coded risk indicators
Rules & Workflow Control
Post-scoring, configurable rules route cases intelligently. High-risk cases go to analysts. Medium-risk cases trigger adaptive authentication. Low-risk cases proceed instantly.
- ✓Threshold changes without deployments
- ✓A/B test models safely in shadow mode
- ✓Automated triage workflow
- ✓One-click rollbacks
Analytics, Audit & Reporting
Complete observability into model performance, fraud patterns, and compliance. Automated drift detection alerts the team to model degradation before it impacts production.
- ✓Daily accuracy and drift reports
- ✓Real-time compliance audits
- ✓Complete decision audit trails
- ✓On-call dashboards with full context
07 — Delivery & Engineering
How We Delivered
Discover
Two weeks of data analysis, workflow auditing, and technical assessment
Output: Architecture roadmap and feature requirements
Define
System design, model selection, and API specification
Output: Architecture diagrams and design documents
Design
Feature store and model pipeline development
Output: Feature engineering pipeline with 200+ signals
Build
Parallel development of API, console, and infrastructure
Output: Production-ready platform with < 50ms inference
Validate and Launch
Shadow mode testing, gradual rollout, and team certification
Output: Production deployment with full monitoring and runbooks
Strategic Decisions
Key Engineering Decisions
Challenge
Achieving sub-200ms latency at 40K+ transactions per day when naive model inference would take 650ms
Decision
Pre-compute expensive features asynchronously into a feature store. Move real-time path to Redis lookup + lightweight scoring.
Outcome
p99 latency dropped from 650ms to 180ms. Checkout unaffected. Feature store in Redis enables < 5ms lookups.
Challenge
Making machine-learning decisions explainable without requiring analysts to understand ML mathematics
Decision
Choose XGBoost with SHAP feature-attribution over a more accurate deep-learning model. Prioritize explainability.
Outcome
Analysts see top contributing features for each decision. Average case resolution dropped from 6 minutes to 90 seconds.
Challenge
Preventing regulatory and compliance violations while unifying signals from multiple PCI-scoped systems
Decision
Build separate PCI-scoped pipelines. Tokenize cardholder data before it leaves transaction systems. Join only anonymized features.
Outcome
Model has unified risk visibility while maintaining PCI-DSS compliance. Complete audit trails.
08 - Impact & Testimonial
The Results
Primary Business Impact
38%
Reduction in fraud losses
Business Metrics
$684K
Annual fraud prevention value
60%
Reduction in false declines
4.0x
Return on platform investment
Engineering Metrics
180ms
p50 decision latency
99.95%
Platform availability
200+
Behavioral features engineered
Before & After
Analyst Review Time
Before
6 minutes per case
After
90 seconds per case
False Positive Rate
Before
15% decline rate
After
2% decline rate
Rule Update Cycle
Before
2 weeks per release
After
30 minutes via shadow mode
“Before Loam, we were trapped in a cycle where tightening fraud rules meant rejecting legitimate customers, and loosening them meant letting fraud through. This platform gave us a completely different approach. Now we're catching fraud while letting legitimate customers transact frictionlessly.”
Sarah Chen
VP of Fraud & Risk
Digital Banking Platform
Related Case Studies
Ready to Transform Your Risk Management?
Whether you're scaling fraud prevention, building real-time intelligence systems, or modernizing legacy platforms, we can help you get there.