Ledgerly Data
Synthetic dataset Synthetic dataset: the shared benchmark
A synthetic B2B SaaS company with CRM, billing, product usage and support data, plus hidden ground truth to score analytics and AI work against. The foundation for most of these projects.
- Status
- Code on GitHub
- Relevant to
- Data and analytics
Results · Synthetic dataset: the shared benchmark
- Synthetic datasetSynthetic dataset: the shared benchmark. A property of the generated data itself, not a real company.ARR at the snapshot, up from $12.7M two years earlierSynthetic dataset
- Synthetic datasetSynthetic dataset: the shared benchmark. A property of the generated data itself, not a real company.active customers: 457 SMB, 208 Mid-Market, 44 EnterpriseSynthetic dataset
- Synthetic datasetSynthetic dataset: the shared benchmark. A property of the generated data itself, not a real company.net revenue retention, trailing 12 monthsSynthetic dataset
- Implementation factA fact about what was built and tested, not a business outcome.invariant and realism testsImplementation fact
The problem
Public datasets rarely look like what a RevOps, analytics or CS team actually works with, and real company data can’t be shared.
What I built
- A simulation of Ledgerly, a fictional seat-based finance-operations platform, from its first marketing lead in January 2023 to a snapshot on September 30, 2026.
- Exports that look like Salesforce, a billing system, product analytics and a support desk, messiness included: duplicate accounts, bad emails, stale close dates, wrong ticket categories.
- The underlying truth (customer health, real duplicates, real ticket categories, how open deals actually end) is exported separately, so models and data-quality tools can be scored rather than just demoed.
- Deterministic output: the same seed gives identical data, and a test checks it.
Stack
Python, DuckDB, Parquet, Faker, pytest