V3I6P39

Roam: An AI-Powered Hyperlocal Discovery and Loyalty Platform for Tier-2 Indian Cities Using Large Language Models and Real-Time Geospatial Intelligence

Ramsha Khan^1*, Dr. Kirti Jain², Ms. Neha Tiwari³

Abstract

India’s Tier-2 urban centres, including Bhopal (1.9 million population) and Kanpur (3.2 million population), collectively present a significant underserved digital market. Despite a combined smartphone penetration rate of 78% across these cities, no locally-intelligent discovery platform or affordable SMB loyalty system has been deployed at scale to serve this population.

Existing national platforms such as Zomato and Google Maps fail to account for local cultural context, Hindi language preferences, neighbourhood-level recommendations, and seasonal behavioural patterns. Furthermore, 91% of small and medium businesses (SMBs) in Tier-2 Indian cities operate with paper loyalty cards or no loyalty infrastructure whatsoever. This paper presents Roam, a mobile-first platform designed to address both dimensions of this gap. Roam integrates: (a) AI-powered local discovery using Anthropic’s Claude API (claude-sonnet-4-20250514) with seven distinct intelligent features, (b) QR-based digital loyalty management for SMBs, and (c) a community social layer with real-time crowd data.

The system integrates the Claude large language model for contextual, bilingual (Hindi/English) recommendations; the Google Maps Places API for verified real-venue data; Supabase PostgreSQL with PostGIS for geospatial queries; and Upstash Redis for caching, supporting scalability to 100,000+ monthly active users.

Evaluation results demonstrate a 97.7% unit test pass rate across 87 tests, Lighthouse Performance scores between 83 and 94 on mobile, and an API p95 response time of 247ms under 1,000 concurrent users. User acceptance testing with 12 participants yielded an average satisfaction score of 4.64 out of 5.0. This work constitutes the first academic documentation of LLM-powered hyperlocal discovery specifically engineered for the Tier-2 Indian cultural and linguistic context, and demonstrates that Claude-class models outperform traditional collaborative filtering on cold-start recommendation by 31–47% in user satisfaction.

Keywords:

Large Language Models, Hyperlocal Discovery, Mobile Loyalty Systems, Tier-2 Cities, Bilingual AI, Google Maps API, Recommendation Systems, Claude API, India, Bhopal, Kanpur