Data pipelines
Global Footprint Mapping
Found and geocoded where suppliers actually manufacture, adding 59,791 verified locations across 56 countries.
Data engineer
Aug 2026
In numbers
- new locations for 9,964 suppliers
- 59,791
- new locations for 9,964 suppliers
- of locations geocoded
- 99.3%
- of locations geocoded
- to research 1,923 headquarters
- ~$81
- to research 1,923 headquarters
The problem
In the buyer app, suppliers showed up with a headquarters and nothing else. Their plants, branches and regional offices weren't in the database, even though where a supplier produces is often what decides the deal.
What I did
- 1
Gather sources
AWS AgentCore's web search ran about three searches per company, 6,419 in total. From the 27,446 links that came back, I put official company sources first and extracted their page content.
- 2
Check identity and role
Amazon Nova on Bedrock checked each address: does it really belong to this company, or to a distributor or subsidiary? Each location also got a role: headquarters, plant, branch or registered office.
- 3
Geocode, never delete
Addresses were geocoded; where the exact address couldn't be found, the location got city-level coordinates instead (26,379 of them). The 427 records without a resolvable city were kept, just without coordinates.
Result
About 60,000 new production sites became visible on the map in the buyer app. Researching 1,923 headquarters cost about $81 in total, roughly 4 cents a company.
What I learned
A location's role matters as much as the address itself. Mix up a plant and a sales office and you send the buyer to the wrong place.
Stack
- Bedrock (Nova)
- AgentCore Web Search
- Python
- PostgreSQL