Stacks/Fintech / internal platform
Stripe Radar's Agentic MLE
Stripe built Agentic MLE so AI agents can carry a fraud-model experiment from idea to result inside its ML infrastructure, with engineers reviewing hypotheses and results. Paired with training that went from 24 to 4.8 hours per epoch on 11 billion transactions, agents implemented and evaluated human-prioritized ideas, and one change lifted Shield's recall by over 8%.
Ryan Drapeau and Sai SamantPublished Oct 19, 2026Last verified Nov 13, 2026
- Use case
- AI agents that run ML experiments on Shield, Radar's transaction-fraud model: investigating missed fraud, implementing candidate features, training and evaluating models against a baseline
- Category
- Fintech / internal platform
- Team size
- not stated
- Orchestrator / framework
- Agentic MLE is a set of tools and instructions that lets agents work inside Stripe's existing ML infrastructure: query data, inspect workflow executions, read experiment results, write feature code, compute historical feature values on Stripe's feature platform, train, and evaluate against a baseline. Engineers review hypotheses and results at defined points; within a run the agent inspects failures, corrects configs, and retries.
- Memory
- Operational knowledge is packaged as reusable platform procedures plus model-specific documentation the agent loads per task; agents are also given results from earlier experiments and investigations.
- Evals
- Every agent-run experiment is evaluated against a baseline model; humans prioritized 38 ideas from agent hypotheses (from model explanations on missed fraud). The winning change (stochastic weight averaging from epoch 0, eight epochs) improved recall by over 8% in a controlled comparison.
- Observability
- not stated
- Cost (stated in writeup)
- Not stated
- Cost per output (stated)
- Not stated in the writeup
- WhatWeRun estimate
WhatWeRun estimate
Not enough info to estimate
The post doesn't name the agents' LLM or give tokens or experiments per month; GPU training cost isn't stated.
Reviewed Nov 13, 2026
- Observability tools
- not stated
- Other tools
- Stripe feature platform (shared offline/online feature definitions)Amazon S3 (streamed training data)GPU training cluster
- Source
- https://stripe.dev/blog/scaling-radar-fighting-fraud-with-agent-assisted-machine-learning
Ryan Drapeau and Sai Samant · Published Oct 19, 2026 · Last verified Nov 13, 2026
Related
