Acquisition cost
Developed and deployed the first model for the top of the customer-acquisition funnel, reducing acquisition cost by 15% while maintaining the approval rate and origination.
NoVerde · 2019–2021
Xper. From datasets to origination decisions.
I built MLflow-based software to connect the modeling process across credit products.
The work behind the tool
One workflow for the modeling lifecycle.
I designed and developed Xper, MLflow-based experimentation software that brought dataset discovery, preparation, feature engineering, tuning and validation into one workflow across credit products.
The aim was to make the process easier to orchestrate and connect model experiments with acquisition decisions.
Retrospective based on my account of Xper. The original implementation and customer data are not published here.
Delivery at NoVerde
Experimentation, acquisition and production.
Xper
Designed an experimentation platform that shortened model building, making it possible to obtain a model and its validations within a few hours.
Delinquency
Implemented more than five credit models, achieving a 35% reduction in delinquency while maintaining the approval rate.
Production model ownership
Responsible for maintenance and deployment of all production credit-origination models.
Reported results from my NoVerde work. These cover separate initiatives; they are not outputs of the demonstration below.
Workflow
From available data to a decision policy.
- 01Discover datasets
- 02Check & clean
- 03Engineer features
- 04Tune models
- 05Simulate originations
Conceptual workflow based on the original tool’s scope. This diagram does not expose its internal architecture.
An experiment, from start to report
Configure. Run. Compare. Inspect.
Explore the orchestration Xper made easier: select a dataset, prepare its features, search model parameters and inspect a recorded experiment.
Why this mattered in my work
Reusable modeling process
Connected data preparation, experimentation and tuning in one cross-product workflow.
Business interpretation
Extended the discussion beyond a model metric to acquisition and origination scenarios.
Credit Suite
The data foundation for MLOps and analytics.
I developed Credit Suite: eligibility, feature, labeling and monitoring services that formed the underlying data model for MLOps and analytics.
Eligibility
Features
Labeling
Monitoring
Conceptual service map based on my account; the original production architecture is not reproduced.
The people behind the work
A chapter at NoVerde.
With colleagues in 2019, celebrating NoVerde’s third anniversary.
