Bharat Dhrishti AI

Today's frontier vision-language models — CLIP, GPT-4V, Gemini — reliably recognise the Eiffel Tower but largely fail on a Durga Puja pandal in Murshidabad, a step-well in Rajasthan, or a chulha-lit kitchen in rural Chhattisgarh. The visual vocabulary of 1.4 billion Indians is nearly absent from AI training corpora. AI tools deployed in India for healthcare screening, crop advisory, and e-governance rely on models that misidentify Indian visual cues. If AI cannot see India, it cannot serve India.

BharatDrishti is a scalable, mobile-first data collection platform enabling contributors across every Indian district to capture, describe, and submit geo-tagged images of culturally rich Indian realities. The platform supports 22 Indian languages, offline draft saving with background sync, voice-to-text description input, and real-time image quality checks — all optimized for budget Android devices on 2G networks. Internal reviewers use a keyboard-shortcut driven review queue with a district coverage choropleth map. Approved datasets export in COCO JSON, CSV, or HuggingFace-ready format for direct use in AI training pipelines.

I structured the PRD around two distinct user types with fundamentally different needs: contributors (field agents, NGO volunteers, rural citizens) who needed a sub-60-second submission flow on low-end devices, and admins who needed district-level coverage visibility and high-throughput review tools. I mapped the full contributor journey across 6 steps — language selection, GPS location confirmation, image capture, description entry, review, and offline handling — and specified edge cases for every failure mode: blur detection thresholds, multilingual profanity filtering, perceptual hash deduplication, and NSFW on-device screening before upload. The MVP was scoped to 17 features with a clear Phase 2 roadmap including ML-based quality scoring, gamification, and HuggingFace auto-push.

Delivered a complete PRD for JoshTalksAI covering the full platform — contributor app, admin dashboard, technical architecture, and success metrics. If BharatDrishti reaches 50% district coverage within 12 months, JoshTalksAI will have produced the largest structured vision-language dataset of Indian cultural contexts ever assembled — surpassing existing India-specific datasets by an order of magnitude. The platform is live at rushankt.github.io/BharatDhristhiAi/