Wise Drop
Inspiration
Wise Drop was born from a simple observation: India’s groundwater crisis is fundamentally an information problem. Farmers and local administrators make high-stakes water decisions without real-time, localized intelligence. We wanted to apply modern AI so that every farmer, village officer, and policymaker can make data-driven choices that preserve water, protect livelihoods, and align with national water goals.
What it does
Wise Drop is an AI-powered groundwater intelligence platform that delivers actionable, policy-aware guidance to three user groups:
- Farmer Agent: personalized irrigation and crop-water guidance in local languages.
- Village Officer Agent: village-level monitoring, automated compliance reports, and SDG 6 tracking.
- Policy Maker Agent: district/state analytics, predictive modeling, and evidence-based recommendations. The platform ingests groundwater, soil, and weather data and returns short, practical recommendations, alerts, and audit-ready reports.
How we built it
- Frontend: React.js with Tailwind CSS for a responsive UI.
- Backend: Python + FastAPI serving APIs and orchestration.
- AI layer: Multi-agent architecture using LLMs with RAG (vector DB) for domain context and NLP for multilingual responses.
- Data stores: PostgreSQL for structured data and a vector DB for embeddings/context.
- Integrations: Weather APIs, soil data sources, and CGWB monitoring inputs via secure APIs.
- Deployment: Cloud-hosted, containerized services and serverless functions for scalability.
- Compliance: Agent-level guardrails implement CGWB and Jal Jeevan Mission alignment and an audit trail for every recommendation.
Challenges we ran into
- Data availability and heterogeneity: Groundwater and soil data come in inconsistent formats and coverage across regions.
- Multilingual accuracy: Ensuring high-quality advice in Tamil, Marathi, Hindi, Telugu, and Kannada required iterative tuning and vernacular prompts.
- Policy guardrails: Encoding CGWB/JJM rules into agents so they refuse unsafe recommendations took careful design and auditable logic.
- Latency vs. accuracy: Balancing model choice and prompt complexity so recommendations are fast enough for real-world use.
Accomplishments that we're proud of
- Built a live platform with three functional AI agents reachable at the public demo URL.
- Aligned the product design to Jal Jeevan Mission, CGWB guidelines, and UN SDG 6 from day one.
- Successfully trimmed and prepared demo material for government and hackathon submissions without vendor-branding.
- Packaged a production-ready README, architecture docs, and a 5–7 slide pitch deck for fast submissions.
- Completed ET AI Hackathon submission (Track 5 — Domain-Specialized Agents) with edge-case demos and compliance guardrails.
What we learned
- Building AI for public-sector impact requires domain-first design: policy alignment, trust, and auditability matter as much as raw model accuracy.
- Iteration with real users (village officers and farmers) is essential — simulated tests miss practical edge cases.
- Multi-agent design simplifies responsibilities: separating farmer, officer, and policymaker concerns made the product both easier to test and more useful.
- Keep the public-facing assets neutral on vendor branding when engaging government stakeholders — focus on impact and policy fit.
What's next
- Pilot deployment in Tamil Nadu with a partner district to collect live usage signals and validate the impact model.
- Deepen integrations with CGWB data and state monitoring systems via secure APIs.
- Improve multilingual capabilities with more vernacular testing and small fine-tuning where permitted.
- Harden auditability and escalation flows so agents can escalate ambiguous or high-risk cases to human experts automatically.
- Prepare funding proposals and CSR partnership packs to scale pilot deployments across multiple states.
Log in or sign up for Devpost to join the conversation.