Identity resolution
CompletedGolden Record
A system that identifies when different records describe the same customer, reconciles them into a single trusted record, and keeps a clear history of every decision.
- Status
- Completed
- Type
- Identity resolution
- Technology
- Node.js
- TypeScript
- PostgreSQL
- Google Cloud
The challenge
Customer data often arrives from several sources with inconsistent spelling, formatting, and completeness. Duplicates inflate counts, split customer history across records, and make downstream reporting unreliable. Any fix has to be accurate, explainable, and safe to run on data belonging to different organizations.
The approach
Golden Record standardizes incoming data, then evaluates candidate pairs in two stages: deterministic rules for high-confidence matches and probabilistic scoring for records that are similar but not identical. Matched records are reconciled into a single golden record while every source record stays traceable.
Engineering highlights
Deterministic matching
Exact rules on normalized identifiers resolve the clear-cut cases quickly and predictably.
Probabilistic matching
Weighted similarity scoring across multiple attributes handles typos, nicknames, and partial records.
Record reconciliation
Matched records are combined into one golden record, with links back to the source records it came from.
Matching architecture
Normalization, candidate selection, scoring, and merging are separate stages that can be tuned and tested independently.
Database integrity
PostgreSQL constraints and transactional updates keep the data consistent as records are merged.
Tenant isolation
Each organization's data is kept separate, so matching never crosses tenant boundaries.
Auditability
Match and merge decisions are recorded, so it is always possible to answer why two records were combined.
Testing
Automated tests cover matching rules and edge cases so changes to the logic can be made with confidence.
Related services
A note on confidentiality
This summary describes the problem and the engineering approach. Source code, data, credentials, and organization-specific details are not shared publicly. We are happy to discuss the technical decisions in more depth in conversation.