Product demonstration
Interactive demo · Synthetic data
One question, followed all the way to a measurable result.
Praxis at Ridgeline Home Services, a fictional company, in seven moments: from connected systems to a second investigation that reuses the first one's foundation.
Lead volume is stable, but collected cash is falling. Why?
The walkthrough
01 · Atlas
Seven systems become one model
Each system keeps its own IDs and field names. Praxis proposes one model of leads, customers, jobs, invoices and payments, and a person approves it. Same-person links wait for review.
- Source systems in one model
- 7

Fabric: one model at revision 2, read by Atlas, Probe and Forge. 02 · Probe
The question branches into explanations
Cash can fall for several reasons that look the same from one tool. Probe writes each one down as a testable explanation and names the systems it needs.
- Explanations tested
- 7

Leads hold steady while collected cash falls 23.8%. Invoicing within 14 days moved most. 03 · Probe
Evidence narrows them
Each explanation is tested on the raw data and again within segments that could confound it. A plausible story can collapse: one salesperson's raw gap of 20.1 points shrinks to 1.9 within segments. A supported one is an association, not a proven cause.
- Supported / weakened / not testable
- 2 / 4 / 1

Raw and adjusted differences with 95% intervals. Two plausible explanations are weakened. 04 · Probe
Opportunities are ranked by value
Each supported finding gets an estimated value range with its assumptions. Revenue, collected cash, contribution and cost are kept apart, and opportunities are ranked by estimated collected cash.
- Top opportunity, estimated collected cash (synthetic data)
- $1.44M to $2.74M

Ranked by estimated collected cash. Ranges are estimates on synthetic data, never money earned. 05 · Forge
An approved change moves into execution
The fix becomes a change set on named records. Rules can block it, a second person approves that exact set, and it runs in a sandbox. Nothing is sent to a real customer.
- Sandbox records applied and read back
- 87 of 87

Proposed, approved and executed by three different people, in the sandbox. One visible exception, safe to retry. 06 · Forge
Execution is verified, then measured
Every change is read back to confirm it happened. Outcomes are compared with a holdout group chosen before the run. Until the data arrives, the readout says waiting.
- Readout status
- Waiting, due 2026-12-08

Treated against a holdout Forge picked before the run. The readout waits for data and claims nothing yet. 07 · Atlas
The second question builds on the first
A finance question reuses the same model, sources, links and evaluator. Praxis counts what was reused and what had to be added.
The second questionWhy are completed jobs taking longer to produce collected cash?
- Model elements and sources reused / new
- 16 / 4

16 reused, 4 new. No new sources, object types or relationships.
Run the same method on your own question.
Start with a read-only check on the systems involved. Nothing in your systems changes, and you decide whether to go further.