Agentic improvement loops for your ML systems

Your agents run their own experiments on Zipline, from the first backtest to production, with guaranteed training-serving consistency.

[d] Schedule a demo

[v] View docs

Chronon, the engine behind Zipline, runs in production at

+8 %

recall on Stripe's fraud model

Even mature models have room to improve

Stripe's agents lifted the recall of Radar, their fraud model, by over 8%. On a model that mature, most experiments shift it by less than 0.1%.

Those experiments run on Shepherd, Stripe's feature platform, which is built on Chronon. We created Chronon at Airbnb, and some of us led the team that built Shepherd.

Read Stripe's write-up

What Zipline gives your agents

An agent can come up with far more feature ideas than a team could ever test by hand. Zipline lets it try each one against history and put the good ones live.

Backtest any idea

Your agent, whether that's Codex, Claude Code or something else, writes a feature in Python. Zipline works out what that feature would have been at every point in the past, with no future data leaking in. The agent then trains a model on it and compares it against your existing baselines.

Put the winners live

The Python the agent wrote becomes the production pipeline. Zipline generates the Flink jobs, the backfills and the online store updates, and the model is served exactly the values it was trained on.

Safe to run many at once

A branch for every experiment

Each experiment gets its own branch. Agents can run lots of them side by side without going near production jobs or tables.

Surgical re-computation

Zipline diffs the statically inferred dependency graphs of the old and new pipelines and reuses everything from production that hasn't changed. That cuts experimentation time and cost by 10x.

Stays in your cloud

Data, code and metadata stay in your own cloud account and never leave it.

Some of the largest payment, underwriting and recommendation systems already run on our engine.

Cheaper to run

On average, customers cut their existing infrastructure costs by 4–7x. In benchmarks, Zipline comes out cheaper overall than the alternatives.

“At Credit Karma, when comparing with BigQuery, we see large improvement on large window backfill. Generally, 7x cheaper for streaming compared to Dataflow and 3x cheaper and faster for large window backfill compared to BigQuery.”

Tran Nguyen

Senior Engineering Manager, Whatnot, ex-Credit Karma

Built for very large systems

Our engine sits behind some of the largest ML systems in the world.

“Chronon allows us to accelerate feature work in the Sora 2 launch. Building a scalable system like this from scratch would have taken us a lot longer!”

Mick Jermsurawong

Member of Technical Staff, OpenAI

Training matches serving

Backfills for training and online features for serving come from the same definition and the same engine.

“Chronon's point-in-time correctness and single engine for both historical and real-time processing let us migrate the vast majority of Stripe's fraud-detection features in just a few months, serving thousands of features to our models while meaningfully cutting infrastructure spend with no impact on latency or availability.”

Arash Rai

ML Engineer, Stripe

End-to-end lineage

Features and signals live in one place, where they're easier to understand, trace and look after.

“Before Chronon the feature pipelines were a disaster. Code was everywhere, some in Java, some in Scala, some in SQL. It was very hard to trace a feature. Now it’s 10 lines of code where before there were 100s of lines across languages.”

Daochen (Frank) Zha

Machine Learning Engineer, Airbnb

Cutting-edge performance

Fresh features at low latency, for models that decide in real time.

“Showing relevant content and fresh content to personalize users' active sessions is an extremely tough challenge for any platform. Zipline solves exactly this problem. High-velocity transaction fraud and sub-millisecond data transformations, aggregations are the need of the hour to find patterns to prevent fraud. Zipline is the best-in-class at these problems.”

Raj Katakam

Staff ML Engineer, Credit Karma

<10 ms

P95 latency

<1 s

P99 streaming freshness

100b+

rows training data scale

1m+

QPS throughput

Built-in monitoring

Drift detection, segmentation and data quality checks come as standard. There's no separate tool to buy and wire up.