You Cannot Build AI on a Data Swamp.
We Clean It First.

The Sovereign Pipeline End to End

Six stages. One machine. No cluster. No cloud database fees.

Stage 1 — The Data Acme Analytics Corp · Series A SaaS

Three years of operational data from a fictional Series A SaaS company — the exact profile of a typical new client engagement.

The grime — three systems that were never designed to talk to each other:

Stage 2 — Data Quality Audit Before Cleaning

Automated quality score across every table — nulls, format inconsistency, duplicate rate. This is the first deliverable clients receive.

Stage 3 — Cleaning SQL

Type coercion, deduplication, date normalization, FK repair, amount parsing — all in a single SQL query per table.

-- Clean transactions: parse amounts, normalize dates, dedup, recalculate totals
CREATE TABLE transactions_clean AS
WITH
  amount_parsed AS (
    SELECT *,
      TRY_CAST(REGEXP_REPLACE(TRIM(line_total), '[\$,USD ]','','g') AS DOUBLE) AS line_total_d,
      TRY_CAST(TRIM(quantity) AS INTEGER) AS quantity_i
    FROM transactions_raw
  ),
  deduped AS (
    SELECT *,
      ROW_NUMBER() OVER (PARTITION BY transaction_id ORDER BY transaction_date) AS rn
    FROM amount_parsed WHERE line_total_d > 0 AND quantity_i > 0
  )
SELECT
  transaction_id,
  TRY_STRPTIME(transaction_date, ['%Y-%m-%d','%m/%d/%Y','%d-%b-%Y','%B %d, %Y']) AS transaction_date,
  -- Recalculate if calc error detected (|reported - computed| > 2%)
  CASE WHEN ABS(line_total_d - quantity_i * unit_price_d) > 0.02 * unit_price_d
    THEN ROUND(quantity_i * unit_price_d, 2)
    ELSE line_total_d
  END AS line_total,
  UPPER(REGEXP_REPLACE(currency, '[^A-Za-z]','','g')) AS currency
FROM deduped WHERE rn = 1
  AND UPPER(TRIM(currency)) IN ('USD','EUR','GBP')

Stage 4 — Analysis 5 Queries · 43 ms Total

Five analytical queries across 962,534 clean transaction rows. This is the benchmark inside the demo.

Anomaly Detection — IQR + Z-score

Stage 5 — AI Document Intelligence RAG · Mistral NeMo 12B

The cleaned data summaries, anomaly reports, and 200 enterprise contracts are automatically ingested into the vector store via ra-watch — no manual steps. Mistral NeMo 12B runs locally on the GCP A100. Zero OpenAI fees. Zero data egress.