Compliance monitoring · Production

Instacart builds real-time fraud detection with Yoda decision platform and ClickHouse

The problem

Instacart needed to detect a wide variety of fraudulent patterns — including fake accounts, payment fraud, and customer-shopper collusion — in real time to protect its financial health and platform trust, while keeping latency low enough for in-transaction decisioning.

Workflow diagram · grounded in source
1
Analyst manages fraud rules
human_review
“Analysts create, read, update, and delete rules through a UI.”
2
Decisioning request arrives
trigger
“A request comes in for decisioning into the real-time system.”
3
Feature fetching incl. ML inference
ai_action
“Feature system fetches and aggregates features based on the request information and the relevant rules configurations. Features can come from machine learning inference services, Instacart's in-house Feature Store, and ClickHouse.”
4
Rules evaluation and decisioning
validation
“Evaluation service evaluates the rules, comparing retrieved features against the rule logic for decisioning and actioning.”
5
Action dispatched to services
output
“Action dispatch consolidates evaluated actions and calls appropriate action services. Examples of actions include shopper warnings, suspensions, and issuing selfie identification checks.”
Reported outcome

Real-time fraud detection with Yoda resulted in millions of dollars a year in savings, and the shift to self-serve analyst-led rule creation reduced engineer time spent on heuristic development from 80% to 20%.

Reported metrics
Annual fraud savingsmillions of dollars a year
Engineer time on heuristic developmentdecreased from 80% to 20%
Reported stack
YodaClickHouseDebezium's Postgres connectorKafkaKafka ConnectFlinkmachine learning inference servicesFeature Store
Source
https://tech.instacart.com/real-time-fraud-detection-with-yoda-and-clickhouse-bd08e9dbe3f4
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Real-time fraud detection with Yoda resulted in millions of dollars a year in savings, and the shift to self-serve analyst-led rule creation reduced engineer time spent on heuristic development from 80% to 20%.

What tools did this team use?

Yoda, ClickHouse, Debezium's Postgres connector, Kafka, Kafka Connect, Flink, machine learning inference services, Feature Store.

What results were reported?

Annual fraud savings: millions of dollars a year; Engineer time on heuristic development: decreased from 80% to 20% (source-reported, not independently verified).

How is this compliance monitoring AI workflow structured?

Analyst manages fraud rules → Decisioning request arrives → Feature fetching incl. ML inference → Rules evaluation and decisioning → Action dispatched to services.

Related compliance monitoring cases