Scope: Detect duplicate identities across noisy, fragmented datasets (fraud, patient mismatch, citizen records)
Architecture: CNN Embeddings + Siamese Network

Problem: Real-world identity data is messy, small inconsistencies cause one person to appear as multiple records, creating operational risk and inefficiency.

Approach: We learn record similarity using CNN embeddings and a Siamese network. LinkID detects, ranks, and resolves duplicate identities auto-linking high-confidence matches and routing borderline cases for review.

Data: HPI snapshot of North Carolina voter records with labeled duplicate and non duplicate pairs.

Results: Strong performance overall, with ~25-point improvement on hard cases where traditional models struggle.

Conclusion: Learned similarity models significantly outperform traditional approaches in complex identity resolution tasks.