Platform

One platform to generate synthetic time-series data - then test where your AI breaks.

A modular platform: use one module or chain them end-to-end - generating the data your evals need and measuring the results, for ML testing and agent evaluation.

rockfish · modules composable
DataFuelsample-based
SchemaFuelschema-only
Scenario Studioscenarios
AgentFuelagent evals
The Rockfish workflow

From input to baseline data - then diverse data to test your AI.

Generate data or train a generative model, measure its quality, then use it to produce the diverse, labeled data your AI testing needs.

1
Input

Data or Schema

Production time-series, a schema definition, or plain-language intent.

2
Rockfish Platform

Train or generate

Take a data sample, train a generative model, and generate baseline data (DataFuel) - or start from a schema or intent and generate it directly (SchemaFuel).

3
Output

Baseline data

Either route lands here - baseline data, ready for downstream use.

Then: diverse data to power your AI testing
Expand your data

Scenario Studio

Scenario Studio

Takes your Rockfish generative model or dataset and produces scenario-specific variations - injecting edge cases, rare events, and incident patterns to expand your training data.

Evaluate your agents

Scenario Studio + AgentFuel

Scenario Studio → AgentFuel → AgentEval

Scenario Studio injects patterns into your data, then AgentFuel generates prompts, queries, and expected responses - ready for AgentEval to score.

The modules

Use one module, or the whole pipeline.

Each module handles one part of the workflow - run a single one on its own, or chain them end-to-end for ML testing and agent evaluation.

DataFuel
Sample-based generation

Generate synthetic time-series from a sample of your production data - learning its structure, correlations, and behavior.

SchemaFuel
Schema-only generation

Generate from a schema definition alone - no source data required to get a realistic, eval-ready dataset.

Scenario Studio
Scenario injection

Inject edge cases, rare events, and incident patterns into the generative model's output or a dataset - to expand your training data.

AgentFuel
Agent eval generation

Turn scenarios into prompts, queries, and expected responses - the eval suites your agents get scored against.

Enterprise-ready

Ready for the way you actually run.

Fits your stack

Outputs land in Snowflake, Databricks, or your existing pipeline - with no model lock-in.

Flexible deployment

Run in your own VPC, on-prem, or as a Snowflake Native App - generation happens where your data lives.

Secure by design

SOC 2 Type II, independently audited. Rockfish learns patterns, not people - raw records never leave your environment.

Built for continuous data

Regenerate as your schema, logic, or models change - so your eval data never goes stale.

Put the platform to work on your data.

Bring a schema or a sample, generate what your evals are missing, and measure the result - end to end.