Finance ops · Production

Stripe builds Shepherd ML feature engineering platform on Chronon to power fraud detection at scale

The problem

Stripe's ML feature engineering infrastructure needed to meet strict requirements for both low latency and high feature freshness across massive transaction volumes, and their existing platform could not adequately scale to these demands.

Workflow diagram · grounded in source
1
Feature definition via API
trigger
“ML engineers use Chronon to define their features with a Python- and SQL-based API”
2
Offline batch computation
integration
“The Chronon offline algorithm produces both offline training data for models and batch-only use cases”
3
Streaming updates via Flink
integration
“With Flink now powering our feature updates, we achieved p99 feature freshness of 150ms”
4
Feature serving from KV store
output
“computing a feature only requires retrieving and aggregating the tiles for the feature rather than all the individual events”
5
Fraud model inference
ai_action
“the Shepherd-enabled model has outperformed our previous model, blocking tens of millions of dollars of additional fraud per year”
6
Consistency monitoring
validation
“we also built monitoring and alerting for Shepherd—including integrating with Chronon's online offline consistency monitoring”
Reported outcome

Shepherd's SEPA fraud detection model, with over 200 features, outperforms the previous model and blocks tens of millions of dollars of additional fraud per year, while streaming feature updates achieve p99 freshness of 150ms.

Reported metrics
Additional fraud blocked per yeartens of millions of dollars
P99 feature freshness150ms
SEPA fraud model feature countover 200
SEPA fraud model performance vs previous modeloutperformed our previous model
Reported stack
ChrononShepherdFlinkAirflowSpark SQLIcebergHive
Source
https://stripe.com/blog/shepherd-how-stripe-adapted-chronon-to-scale-ml-feature-development
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Shepherd's SEPA fraud detection model, with over 200 features, outperforms the previous model and blocks tens of millions of dollars of additional fraud per year, while streaming feature updates achieve p99 freshness…

What tools did this team use?

Chronon, Shepherd, Flink, Airflow, Spark SQL, Iceberg, Hive.

What results were reported?

Additional fraud blocked per year: tens of millions of dollars; P99 feature freshness: 150ms; SEPA fraud model feature count: over 200; SEPA fraud model performance vs previous model: outperformed our previous model (source-reported, not independently verified).

How is this finance ops AI workflow structured?

Feature definition via API → Offline batch computation → Streaming updates via Flink → Feature serving from KV store → Fraud model inference → Consistency monitoring.