← back to index
01WORK

Heera.

software intern · jul 2025 — oct 2025 · pune

Records scored
~13,500
Components
Google Maps · fuzzy · reverse-geo cosine
Clients
logistics incl. Red Bull
Status
shipped to production data pipeline

Problem

Existing internal datasets contained address strings of mixed quality — partial matches, transliteration variance, missing localities. The team needed a scoring service that validated each record and emitted a confidence score plus the component contributions, so downstream pipelines could threshold or route by confidence.

Approach

Three signals, combined into a weighted score:

  1. Google Maps API match — query the address; compare returned formatted address against input.
  2. Fuzzy string match — token sort + ratio over normalized strings to catch transliteration / spacing variance.
  3. Reverse-geocoded cosine similarity — geocode the address, reverse-geocode the resulting lat/lng, compare embeddings of the two textual addresses.

The combined score and per-component breakdown were attached to each record. Approximately 13,500 records were scored during the engagement.

Impact

The service gave downstream teams a rapid, repeatable way to trust an address before acting on it — improving address confidence for logistics clients including Red Bull. It’s a scoring service, not a classifier: it emits a confidence score and its component contributions, and leaves thresholding to the consumer.