Field guides

Understanding production-grade synthetic data

Short, concrete guides to what separates data you can ship on from data that only looks right. Written for the engineer who has been burned by fake data before.

Start here
How-to · 5 min read

Generate your first dataset from a template

A five-step walkthrough with screenshots: pick a prebuilt industry schema, review the plan, generate straight to S3, and read the scored quality report.

Read the guide →
How-to · 5 min read

The dataset that started as one sentence

No template fit, and writing DDL for a throwaway fixture felt like a bad trade. So he described it in a sentence and let CrossRow draft the plan.

Read the guide →
Concepts
Concept · 6 min read

What “production-grade” synthetic data actually means

Realistic-looking values are easy. Data whose ledgers reconcile, foreign keys resolve, and state machines walk legal paths is a different problem. Here is the bar, with worked examples.

Read the guide →
Problem · 5 min read

Why random fake data breaks your integration tests

Faker gives you well-formatted nonsense. The moment your tests assert anything about relationships between rows, that nonsense starts failing in ways that look like your code’s fault.

Read the guide →
Concept · 7 min read

The demo that kept falling apart

An engineer builds a demo dataset the quick way and watches it break one invariant at a time. Each failure is a lesson in the rules that only live across rows.

Read the guide →
Trust · 5 min read

How CrossRow scores its own output

Every run ends with an independent verification pass and two scores. Here is what Structure and Realism measure, and why a generator that grades itself is one you can put in front of an auditor.

Read the guide →
For ML & LLM teams
Machine learning · 7 min read

Synthetic data for fine-tuning: teaching models the right patterns

Text-to-SQL, structured extraction, and tabular training. Three fine-tuning use cases, and which capabilities keep a model learning real relationships instead of a generator’s artifacts.

Read the guide →
Machine learning · 6 min read

The fraud model that aced the test and failed in production

A detector scores 0.98 in validation, then misses real fraud. The training anomalies had the wrong shape. Why real anomalies cluster, and how to generate them that way.

Read the guide →
How-to · 6 min read

When your churn model thinks everyone is leaving

Train on a balanced synthetic set and you ship a model that cries churn at every customer. How to generate the imbalance you actually face, with exact labels.

Read the guide →
Machine learning · 6 min read

The text-to-SQL model that wrote perfect queries against nonsense

The SQL parsed and executed. The answers were still wrong, because the database underneath did not hold together. The substrate that fixes it.

Read the guide →
Workflows
Workflow · 6 min read

The wallet demo that wouldn’t balance

A finance reviewer catches that the ledger doesn’t foot. The rebuild takes one plain-English sentence. A workflow for chained-balance data that reconciles.

Read the guide →

See it on your own schema

Generate a scored sample in minutes. No production data required.

Open CrossRow