Start with Overview, then click Open Interactive App Prototype, choose either Contributor or Admin Reviewer, verify with OTP, and explore the role-specific experience.
Models like CLIP and GPT-4V can identify the Eiffel Tower instantly โ but fail on a Durga Puja pandal, a handpump in rural Rohtak, or a chulha-lit village kitchen. India's 1.4 billion people are invisible to AI.
BharatDrishti mobilises field agents, NGO volunteers, and individual citizens to capture 1,000 geo-tagged, described images per village. Every submission is reviewed, verified, and packaged into training-ready datasets.
The dataset focuses on visual realities that mainstream global training sets often miss.
Every submission is checked by reviewers so quality, caption richness, and location fidelity remain trustworthy.
The flow is built for rural contribution at scale, including multilingual use and low-connectivity operation.
Drafts auto-saved locally. Background sync when internet returns.
GPS override with manual confirmation. Admin can reassign post-submission.
Inline nudge enforces 10-word minimum. Voice-to-text lowers barrier.
Laplacian blur detection. NSFW on-device screen. Perceptual hash dedup.
Village life, local markets, festivals, farming tools, and regional living spaces are often missing from training data. That makes AI systems less useful for India.
Contributors collect geo-tagged photos and descriptions from villages. Admin reviewers validate quality, location, and captions before the data enters the training pipeline.
For field agents, NGO volunteers, and citizens contributing image + description pairs from the ground.
For internal reviewers validating relevance, image quality, caption richness, and district coverage.
The first screen clearly explains why BharatDrishti exists and how contributors plus reviewers solve the training-data gap together.
Users choose between Contributor and Admin Reviewer before authentication, so the app knows which experience to load next.
A single mobile-number flow handles both roles. OTP verification is the shared trust checkpoint before routing the user.
Contributors go to location confirmation, then camera, description, review, and offline-safe submission.
Admin Reviewers land on a routed access screen that opens the review queue and dashboard instead of the contributor capture flow.
The prototype now demonstrates the exact fork you wanted: same OTP gate, different post-login journeys.
| District | State | Progress | Images |
|---|---|---|---|
| Muzaffarpur | Bihar | 73% | |
| Jaipur | Rajasthan | 61% | |
| Varanasi | UP | 58% | |
| Madurai | Tamil Nadu | 44% | |
| Purulia | West Bengal | 18% |
| Feature | Status | Purpose |
|---|---|---|
| Phone OTP / Google login | MVP | Fast sign-in for contributors and reviewers |
| 22-language UI | MVP | Lower friction for multilingual participation |
| GPS district auto-detection | MVP | Anchor submissions to the right geography |
| Camera + gallery upload | MVP | Support both live capture and existing photos |
| Blur / resolution quality check | MVP | Improve training-data usefulness before review |
| Text + voice-to-text description | MVP | Make rich captions easier to submit |
| Offline drafts + background sync | MVP | Work reliably in low-connectivity regions |
| Admin review queue (A/R/F) | MVP | Quality control for every image-caption pair |
| Coverage choropleth map | MVP | Track district and village progress visually |
| CSV / COCO JSON export | MVP | Package verified data for downstream ML use |
| Feature | Status | Purpose |
|---|---|---|
| ML auto-quality scoring | Phase 2 | Prioritize high-confidence submissions automatically |
| Gamification & leaderboards | Phase 2 | Improve contributor retention and momentum |
| HuggingFace auto-push | Phase 2 | Publish verified datasets faster |
| Video collection support | Phase 2 | Expand beyond still-image datasets |
| Contributor incentives / payments | Phase 2 | Support sustained field participation |
| Public dataset explorer | Phase 2 | Make discovery and transparency easier |
| Auto-captioning pipeline (GPT-4V) | Phase 2 | Speed up review and annotation workflows |
| RLHF annotation interface | Phase 2 | Support higher-order feedback collection |
| Real-time collaboration for reviewers | Phase 2 | Coordinate teams on large review batches |
| Metric | Dimension | Target | Measurement |
|---|---|---|---|
| Districts with โฅ 100 images | Coverage | 800 / 800 | Coverage dashboard |
| Total approved image-caption pairs | Volume | 10M+ | Admin analytics DB |
| Submission approval rate | Quality | โฅ 85% | Reviewer decisions |
| Average description length | Quality | โฅ 25 words | NLP pipeline |
| Non-English multilingual submissions | Inclusivity | โฅ 40% | Language detection |
| Submission-to-review turnaround | Speed | < 48h | Timestamp delta |
| Offline sync success rate | Reliability | โฅ 98% | Background sync logs |
| Contributor month-2 retention | Engagement | โฅ 30% | Cohort analytics |
| App crash rate | Stability | < 0.5% | Sentry reporting |