fellegi-sunter

Zero-config entity resolution feeding a durable identity layer: messy records from any source become stable golden entities, a Customer 360 with provenance, merge/split and audit. Fellegi-Sunter beats hand-tuned Splink. Arrow-native/Rust, 250M rows in 11.2 min. Python + edge TypeScript (WASM), SQL-native in Postgres & DuckDB, 97 MCP tools + REST.

Payout-to-bank-statement reconciliation, scored against a sealed answer key. Every match carries its evidence; every gap carries its reason.

Probabilistic identity resolution into an Apache Iceberg lakehouse - Fellegi-Sunter record linkage (Splink) producing deduplicated golden records

Finds records that describe the same person across messy databases, using a Fellegi-Sunter model that needs no labelled examples. Pairwise F1 0.9909 on 199,684 synthetic records with known ground truth. Sends uncertain merges to human review, and attacks its own privacy-preserving linkage mode.

Record linkage over four disagreeing provider rosters: entity resolution with field-level lineage, per-field source priority, and a human override layer. The matching threshold is a cost decision, not a tuning parameter -- what a false match costs against a missed one, and whether an override survives the next file.

Эталонная реализация разрешения тождества по разнородным выгрузкам поставщиков: нормализация, вероятностный матчинг Fellegi–Sunter, провенанс по полям, воспроизводимый каталог. TypeScript · Node 24 · PostgreSQL 18 · Drizzle · pg-boss. Портфолио-проект, данные вымышленные.

The phone agent that refuses to guess. One disclosed verification call per healthcare directory listing: every answer cites its verbatim transcript span, reconciliation shows its arithmetic, and a conformal gate abstains rather than guess. CALL-E hackathon, Sep 2026.

Probabilistic record-linkage tool that finds and merges duplicate adverse drug reaction (ADR) reports, with a human-in-the-loop review UI