Supplier Atlas
The same factory existed nine times across our systems under nine spellings. Atlas resolves supplier identity across sources and keeps a single record with its history intact, so "have we worked with them before" has an answer.
- Year
- 2024
- Role
- Build
- Built with
- Python, Postgres, pgvector
- 9→1
- records per real supplier
- 2.3k
- entities resolved
The problem
Supplier names are entered by humans in a hurry, transliterated inconsistently, and change when a company restructures. Exact matching finds nothing; fuzzy matching alone merges two genuinely different factories in the same industrial park, which is worse than doing nothing.
Approach
Blocking on strong signals first — registration numbers, addresses, contact domains — then embeddings for the remainder, then a human queue for anything under the threshold. Merges are reversible and audited, because an unreviewable merge in a system of record is a liability.
Status
Running against production data with the review queue staffed. The open question is decay: records go stale and I do not yet have a good signal for when a resolved identity should be re-examined.