Fraud Detection Model, From Notebook to Production

IDS rebuilt the fraud pipeline on governed transaction data, then operationalised the model using Databricks and MLflow for experiment tracking, versioning, and automated retraining, with Azure ML handling deployment and monitoring for drift.

Client

A digital-first retail bank

Services provided

About the project

Our solution

The Customer: Who they are

A digital-first retail bank processing high transaction volumes with a data science team under pressure to reduce fraud losses without frustrating genuine customers.

The Challenge: Identifying the villain

The bank’s data science team had built a promising fraud model in notebooks, but it never made it to production — there was no MLOps infrastructure to deploy, monitor, or retrain it, and the underlying transaction data had quality issues the model quietly absorbed. “Our model worked great in the notebook and never once in the real world,” said the Head of AI/ML.

The Journey: The search for a solution

The team tried to hand the model to the engineering group for manual deployment, but without a repeatable pipeline, every retraining cycle became a bespoke project that ate weeks of engineering time.

The Discovery: Finding IDS

The bank’s CDO connected with IDS through the bank’s existing data governance engagement, recognizing the fraud model’s data quality problems traced back to the same underlying issues.

The Solution: The hero arrives

IDS rebuilt the fraud pipeline on governed transaction data, then operationalised the model using Databricks and MLflow for experiment tracking, versioning, and automated retraining, with Azure ML handling deployment and monitoring for drift.

The Implementation: The battle

The model was rolled out in shadow mode first — scoring live transactions without acting on them — so the fraud team could validate performance before it started blocking transactions. The main obstacle was false-positive tuning; IDS worked iteratively with the fraud operations team to calibrate thresholds against real cost of both fraud losses and customer friction.

The Results: The happy ending

  • False positive rate: reduced by 35% 
  • Fraud caught pre-transaction: +25% vs. rules-based system 
  • Model retraining cycle: manual/ad hoc → automated monthly 
  • Time from model update to production: weeks → days  [placeholder]

“This is the first model that actually survived contact with production.” — Head of AI/ML

Technologies used: Databricks · MLflow · Python · Scikit-learn · Azure ML

Facebook
LinkedIn
Send on E-mail

More use cases

One Platform, Two Data Worlds: A Big Data Analytics Mart for Banking Group Controlling

Client: Leading Banking Group
IDS designed and built a new analytical data mart purpose-built for Controlling, constructed entirely on big data technologies rather than the client’s Oracle stack.

Migrating Without Missing a Beat: A New Data Warehouse Foundation for a Banking Controlling Team

Client: Leading Banking Institution
When the migration to the new DWH began, IDS took on the task of carrying that Datamart — and the business logic behind it — onto the new platform.

From Hadoop to Databricks: Scaling a Banking Data Platform Beyond Its Limits

Client: Banking Client
IDS Consulting was brought in to provide hands-on technical support and consultancy for the migration to Databricks.

From Days to Minutes: An 8-Year Data Warehouse Transformation for a Leading Telecom Operator

Client: Leading Telecom Operator
IDS Consulting was brought in not for a single fix, but as an ongoing extension of the client’s Data Warehouse team. The mandate: modernize how data moved through the platform, and keep it evolving safely as the business — and the underlying infrastructure — kept changing around it.

Full-Stack Data Platform with Embedded Testing & Knowledge Transfer

Client: A mid-size IT services company
IDS delivered the platform end-to-end (Python and .NET services, automated test coverage from day one) while running a structured knowledge transfer track in parallel — pairing IDS engineers with the client’s internal team throughout, not just in a handover week at the end.

A Full-Stack Commerce Data Platform, Built and Tested End-to-End

Client: A Full-Stack Commerce Data Platform, Built and Tested End-to-End
IDS delivered full-stack development (Java backend, React front end) with a dedicated QA workstream running in parallel from the start — functional, regression, and performance testing built into every sprint — under a single programme manager accountable for scope, timeline, and quality together.