Atlas Bank
Real-time fraud detection that saved $11M in 12 months.
Replaced a rule-based fraud system with a custom ML pipeline that scores 50M+ transactions a day in under 100ms.
- 94%Fraud reduction
- $11.2MAnnual savings
- <100msScoring latency
- 78%False-positives reduced
Where they were when we started.
Atlas Bank was losing roughly $12M annually to card-present and card-not-present fraud. Their legacy rule-based system had drifted into thousands of overlapping rules — generating so many false positives that legitimate customers were being blocked at the rate of one in every twelve transactions. Customer NPS was bleeding, and the fraud-ops team was running on tribal knowledge.
How we shipped it.
- 01
Audited 24 months of transaction data and reverse-engineered the rule graph to identify which rules were actually pulling weight versus which were dead code costing latency.
- 02
Designed and trained a gradient-boosted model on 50M+ labelled transactions, enriched with device, behavioral, and graph signals — with a separate adversarial-example pipeline to keep the model robust against drift.
- 03
Built a real-time scoring service in Go with sub-100ms p99 latency, deployed on Kubernetes with multi-region failover and shadow-traffic rollout.
- 04
Replaced the static rule console with a fraud-ops UI that lets analysts label disputed cases, see model reasoning, and push retraining triggers without engineering involvement.
What changed for the business.
- $11.2M in annual fraud-loss savings, validated by independent audit at month 12.
- False-positive rate dropped 78%, recovering an estimated $4.5M in formerly-blocked legitimate revenue.
- Model retraining cadence moved from quarterly to weekly, with a regression eval suite that fails CI on degradation.
- SOC 2 + RBI audit pass on first review thanks to comprehensive lineage + decision logging.
“Big Bold delivered a fraud detection system that exceeded every benchmark we set. ROI was apparent inside the first month, and they passed our security review faster than any vendor we've worked with.”
LogisticsA logistics ops platform that cut deploy cycles from 4 hours to 8 minutes.
Read case study
E-commerceHeadless commerce migration that lifted revenue 40% in six months.
Read case study
HealthcareA unified patient portal serving 2M+ patients across 150 facilities.
Read case studyHave a project in mind?
Let’s build it.
Tell us what you’re working on. We’ll come back within 48 hours with a sharp first take, a rough scope, and a path to v1.