Verifying Recruitment Offers for Nepali Migrant Workers: A Provenance-Aware, Bilingual Decision Support System

Abstract

Prospective migrant workers in Nepal must often
decide whether to pay using fragmented bilingual advertisements
and intermittently available official records. We present Parakh, a
provenance-aware bilingual decision-support layer that combines
Department of Foreign Employment records, declarative redflag
rules, OCR-assisted field extraction, Media and Information
Literacy questions, and grievance escalation. It exposes source
state and missing evidence rather than issuing an official certificate
or autonomous fraud verdict. On 216 synthetic advertisements,
browser Tesseract, CLI Tesseract, and clean parsing achieved
macro field exact match of 80.3%, 82.2%, and 98.8%; completerecord
exact match was 22.2%, 23.1%, and 93.1%. Browser
agency-name exact match was only 21.6%, while browser lotnumber
exact match was 86.0%, with 0.0% observed hallucination
for all methods. A preliminary browser check on 43 scraped
images found correct lot-number extraction in 38/43 (88.4%)
and destination extraction in 35/43 (81.4%). These measurements
provide reproducible software-level and collection-level baselines,
not independently audited natural-ad accuracy or evidence of
improved user decisions.