Ferhat RudvanoğullarıData Engineer
← All work

Data quality & master data

Evidence-Based Data Audit

Cross-checked 23,443 facts about 140 major automotive suppliers against live web sources, with zero failed runs, for $0.37 a company.

Data engineer

Aug–Sep 2026

In numbers

companies audited, zero failures
140/140
companies audited, zero failures
facts checked against live sources
23,443
facts checked against live sources
of profiles had something to fix
71%
of profiles had something to fix
total, $0.37 per company
$51
total, $0.37 per company

The problem

Ahead of a demo for a major automotive customer, the profiles of its most important suppliers had to be right. Founding year, headcount, headquarters: every fact needed a source behind it.

What I did

  1. 1

    Move the pipeline to AWS

    I moved the audit pipeline to AWS. AgentCore's web search finds the sources and Amazon Nova on Bedrock checks every fact against them.

  2. 2

    Budget per call type

    Reasoning and token limits are set per type of call: a simple check gets a small budget, a hard comparison gets more. That helped keep the cost down.

  3. 3

    No source, no fix

    Every suggested correction is stored with its source and never written straight to the database. The results ship with a gold test set, negative controls to catch false positives, and a line-by-line cost breakdown.

Result

All 140 companies were audited without a single failure, 71% of profiles had something to correct, and the whole run cost $51. In a similar pilot, legal-name coverage for 1,119 suppliers went from 58% to 97% for $9.50 in total, under a cent per supplier.

What I learned

A correction without a source isn't a correction, it's a guess. Attaching evidence to every suggestion made review much faster.

Stack

  • Bedrock (Nova)
  • AgentCore Web Search
  • Python
  • PostgreSQL