synthetic cfo
Cold Start Engine
Checking your session…

Forensic-grade ERP data, generated from accounting rules.

synthetic cfo builds a full year of SAP or Oracle books - purchases, sales, cash, journals, payroll - with fraud planted and labelled, so you get an exact answer key no real dataset can give you. No real company data is ever used.

Try the demo Request access
Have an invite? Sign in
No real data, ever
Everything is generated from double-entry accounting rules, from scratch. There is no source data, so there is nothing to leak and no privacy, GDPR or re-identification risk.
An exact answer key
Every planted fraud is labelled by record, control and vector. You can score a detector's precision and recall against ground truth - the one thing real data can never give you.
Reproducible to the byte
The same seed and the same configuration rebuild the identical dataset every time, so any result can be checked by regenerating it.
Real ERP architecture
SAP ECC, SAP S/4HANA and Oracle Cloud table structures, seven modules, 25+ fraud vectors, with double-entry holding on every line and full referential integrity.
Every dataset ships as a four-layer package
Executive brief (whitepaper) Audit evidence + the answer key A reconciled SQLite database Source workbooks mirroring the ERP tables
Built by a chartered accountant who audited and worked inside SAP and Oracle environments at Big 4 firms and global enterprises. That practitioner knowledge is encoded as deterministic generation rules - no language model ever touches the numbers.
Try the demo Request access

Sign in

Access is restricted to invited users.
Need access? Request it here
Privacy·Terms·Cookies
Online

What should the engine build?

Describe the financial environment you need, then refine it in conversation. Built from accounting rules, never from real data.

Deterministic engine
How is fraud embedded?
How do I score a model?
Pharma SAP environment, heavy fraud, 25,000 rows
Oracle dataset for a bank with lapping and kiting
S/4HANA migration test data, light fraud
What will I receive?
Is this reproducible?
Why synthetic data?
Deterministic engine. Your words map to a configuration you confirm. The financial data never touches a language model.

Advanced configuration

Set every parameter explicitly.

01 - ERP Platform
SAP ECC
ECC 6.0. BKPF, BSEG, EKKO, VBAK.
SAP ECC 6.0
SAP S/4HANA
Universal Journal. ACDOCA.
SAP S/4HANA
Oracle Cloud
Fusion Financials. AP, AR, CE, GL, FA.
Oracle Cloud
Migration Pair
Same business, two schemas. ECC + S/4HANA from one seed.
ECC + S/4HANA
Intercompany Group
A holding company: SAP subsidiary + Oracle subsidiary, with elimination fraud across both.
SAP + Oracle
02 - Modules
P2P
Procure-to-Pay
O2C
Order-to-Cash
CE
Cash Mgmt
R2R / GL
General Ledger
HR / Payroll
Insider threat
MM / Inventory
Costing fraud
FA / Fixed Assets
PP&E schemes
03 - Transaction Volume
POC
500 rows
Standard
5,000 rows
Professional
25,000 rows
Enterprise
100,000 rows
04 - Industry
Tech, Media & Telecom
SaaS, IT hardware, services.
Pharma & Life Sciences
CROs, lab equipment.
Manufacturing
Raw materials, equipment.
Financial Services
Trading, infrastructure.
Retail & Consumer
Merchandise, seasonal.
05 - Fraud Intensity
None
Pristine baseline.
0%
Low
Subtle anomalies.
<1%
Medium
Standard training.
~3%
High
Stress-test density.
>8%
06 - Optional
Fiscal Year
Custom rows (max 1,000,000)
Seed (optional)
Fiscal years (fraud arc)
Operational noise (entropy)
Jurisdiction
Blind track (independent fraud)
Foreign currency exposure
Company size

Your jobs

No jobs yet.

Dataset insights

Live analytics computed straight from your generated databases.

Choose a dataset
Pick a completed dataset above.

Benchmark: how good is your fraud-detection AI?

Score a fraud detector, objectively

Think of it as marking an exam. We plant fraud in known places and keep the answer key, so unlike real data, where nobody knows the true answer, we can grade any fraud-detection tool on exactly what it caught, what it missed, and where it raised a false alarm.

Dataset
See it in action

No tool of your own? Watch a fraud detector get graded on this dataset, in one click. Each is a different quality of detector, so you can see the whole range:

Simulated detectors, shown so you can see how the grading works. See a perfect score → (what submitting the answer key itself looks like).
Have your own fraud-detection tool? Grade it

Run your tool (a script, an AI model, audit software) against this dataset, then bring back the record IDs it flagged, and we find them in whatever it exported.

📁
Drop your results file here, or click to browse
CSV or TXT, exactly as your tool exported it.
Optional: only count certain fraud types:

How to use synthetic cfo

From first sentence to scored model in four steps.

Quick start
1
Describe the environment in the Generate tab, in your own words. Platform, industry, fraud level, size - any order, any phrasing. Try: pharma SAP environment, heavy fraud, 25,000 rows
2
Refine in conversation: make it smaller switch to Oracle only P2P more fraud, then say generate
3
Download the four-layer package from the job card or History: whitepaper, audit evidence with ground-truth labels, SQL database, source workbooks. Open Insights on any completed job for live charts.
4
Score your model in the Benchmark tab: paste the record IDs it flagged, get precision, recall, F1 and a grade against the exact answer key.
Special builds

Migration pair: migration pair for a bank gives the same business in both SAP ECC and S/4HANA from one seed.

Multi-year fraud arc: retail company over 3 years, heavy fraud builds consecutive years where the schemes escalate, plus a consolidated database.

Clean baseline: completely clean pharma data, no fraud produces a pristine zero-fraud control environment.

Reproducibility: every job records a seed; the same seed and configuration regenerate a byte-identical dataset. Set a fixed seed under Advanced.

Frequently asked questions
Loading...

Live Twin

A synthetic company's ERP, posting documents in real time. Purchase orders, invoices, payments and journals stream past as the year advances - and some are fraud, hidden in the flow like they would be in a real system. The answer key stays sealed until each scheme's reveal point, so a detector is scored on catching it live, never on peeking. Freeze the world into a downloadable audit package anytime; it matches exactly what you watched.

1Set up a company and press play. It runs the same deterministic engine as the batch product.
2Watch the desk. Live figures move as documents post; fraud rides along unflagged.
3Schemes surface as the clock passes each reveal point. Seal the world into an audit package when done.
Set up the world
-

Document desk

Waiting for the first document...

Schemes surfaced released by world-time

Nothing surfaced yet. Fraud is hidden in the stream until its reveal point.
Sealing runs the same engine to completion and hands you the standard four-layer package. Same seed, so the answer key matches exactly what streamed.

What you receive

Every dataset ships as a four-layer package

Loading...

Dataset insights

Computed live from your generated database

Loading...

Your jobs

Running now and recently finished. Stays here after a refresh.

Loading...

Chats

Your saved conversations

Loading...