PROJECT WALKTHROUGH / INDEPENDENT SYNTHETIC RECONSTRUCTION
Beauty data & AI platformFrom source records to trusted business answers.
A customer needed to visualize how SAP ERP data, governed data products, shared metrics and document context could support useful AI consumption.
01 / ORIGINAL ENGAGEMENT
What I owned.
I built a tailored end-to-end Databricks demo to show the vision and secure customer buy-in: source ingestion, medallion layers, cashflow/RGM products, metric views, document retrieval and agent consumption.
After kickoff, the account generated $3.2M+ in cumulative services billing, alongside substantial unquantified Databricks consumption. Account billing is separate from GTM team closed business and from this reconstruction's outputs.
02 / HOW THE SOLUTION WORKS
Follow the flow.
Explore the interactive architecture and connected components ↗
03 / RECORDED LOCAL EXECUTION
See the example in motion.
Captioned recording rendered from synthetic evaluation outputs. The website presents evidence; run the downloaded project locally to execute its API.
Example scenario
Synthetic invoice, line and receipt records are reconciled before cashflow/RGM metrics are published. Scoped retrieval adds document context to factual answers.
Inspect recorded output +
{
"scenario": "clean",
"publish_version": 1,
"source_hash": "e6755df59326966fbebc919b5513718c4e132d88edc117de20628d32c602148c",
"quality": [],
"bronze_batches": 1,
"trace": {
"request_id": "7b5a56b425884b89b2ab8509faf48b96",
"stages": [
{
"component": "bronze",
"label": "Receive SAP-shaped synthetic batch",
"duration_ms": 0.488
},
{
"component": "quality",
"label": "Validate keys, currency and ledger reconciliation",
"duration_ms": 0.483
},
{
"component": "gold",
"label": "Publish reconciled Silver and Gold snapshot atomically",
"duration_ms": 0.943
}
],
"total_ms": 2.517
},
"run_id": "cd791bb4c35f411a921dd7c3c205bbc2",
"revision": 2,
"role": "finance",
"markets": [
"US",
"GB",
"CA"
],
"metrics": [
{
"metric": "net_revenue",
"value": 29757.01,
"unit": "USD",
"definition": "Net invoiced sales after returns and discounts",
"markets": [
"US",
"GB",
"CA"
],
"rows": 120,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT SUM(net_cents) AS value,COUNT(*) AS rows FROM rgm WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
},
{
"metric": "margin_pct",
"value": 53.043,
"unit": "%",
"definition": "100 × total profit / total net revenue",
"markets": [
"US",
"GB",
"CA"
],
"rows": 120,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT 100.0*SUM(profit_cents)/NULLIF(SUM(net_cents),0) AS value,COUNT(*) AS rows FROM rgm WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
},
{
"metric": "discount_pct",
"value": 6.2093,
"unit": "%",
"definition": "100 × total discounts / gross sales after returns",
"markets": [
"US",
"GB",
"CA"
],
"rows": 120,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT 100.0*SUM(discount_cents)/NULLIF(SUM(gross_cents),0) AS value,COUNT(*) AS rows FROM rgm WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
},
{
"metric": "cash_collected",
"value": 14941.68,
"unit": "USD",
"definition": "Actual receipts applied to invoices as of the demo clock",
"markets": [
"US",
"GB",
"CA"
],
"rows": 60,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT SUM(paid_cents) AS value,COUNT(*) AS rows FROM cashflow WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
},
{
"metric": "receivables",
"value": 14815.33,
"unit": "USD",
"definition": "Invoice balance minus applied receipts",
"markets": [
"US",
"GB",
"CA"
],
"rows": 60,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT SUM(outstanding_cents) AS value,COUNT(*) AS rows FROM cashflow WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
},
{
"metric": "overdue",
"value": 5995.37,
"unit": "USD",
"definition": "Outstanding balances with due date before the demo clock",
"markets": [
"US",
"GB",
"CA"
],
"rows": 60,
"publish_version": 1,
"semantic_version": "metrics-1.0",
"sql": "SELECT SUM(overdue_cents) AS value,COUNT(*) AS rows FROM cashflow WHERE market IN (?,?,?)",
"parameters": [
"US",
"GB",
"CA"
],
"currency": "USD",
"as_of": "2026-10-12"
}
],
"synthetic": true,
"as_of": "2026-10-12",
"counts": {
"invoice_lines": 120,
"invoices": 60,
"documents": 4
},
"retrieval_mode": "Local lexical count vectors; cloud semantic embeddings are a reference"
}Local scenario checks
| Scenario | Result | p50 ms | p95 ms |
|---|---|---|---|
| reconciliation | PASS | 6.142 | 7.118 |
| ratios | PASS | 5.683 | 6.608 |
| bad batch | PASS | 7.601 | 8.618 |
| duplicate | PASS | 7.114 | 8.251 |
| currency | PASS | 7.27 | 7.571 |
| authorization | PASS | 5.271 | 7.81 |
| retrieval acl | PASS | 5.393 | 7.452 |
| clarification | PASS | 5.172 | 7.344 |
Ten sequential fresh-workspace runs per case. Timing includes workspace creation and scenario execution; nearest-rank p95 is the maximum of ten. Local synthetic measurement only. Local synthetic measurements only; no production SLA or cloud performance is claimed.
04 / ENGINEERING EVIDENCE
Go deeper.
Independent example code and data are available on GitHub and in the downloadable source archive. Original client systems, code and internal documents are excluded.