Mausam gpt

Conversational weather intelligence for every Indian — verified, multilingual, actionable.

Comment

Inspiration

Over 600 million Indians — nearly half the country — depend on a livelihood shaped by weather they often can't access or act on in time, including 130 million cultivators. At the same time, 19% of Indians (age 7+) can't read a forecast bulletin at all, and nearly 38% still have no active internet access. India loses an estimated ₹50,000 crore annually to extreme weather in agriculture alone.

The problem isn't a lack of weather data — IMD, ISRO, and global models already produce accurate forecasts. The problem is that this data never reaches the people who need it most, in a form they can actually use. We realized that we can't control the weather, but we can close the information gap that turns a preventable loss into an unavoidable one. That became the seed for MausamGPT.

What it does

MausamGPT is a multilingual, conversational weather intelligence platform that lets anyone — a farmer, a fisherman, a pilot, a disaster management officer — simply ask a weather question in their own language (voice or text) and get a verified, actionable answer back.

Instead of raw numbers, a farmer in Bhopal asking "कल बारिश होगी?" gets a response like: "Yes, 82% chance of rain tomorrow after 3 PM. Avoid pesticide spraying. Irrigation recommended before noon."

It supports 22+ Indian languages via Bhashini, works over Web, WhatsApp, and IVR (so it isn't locked behind a smartphone app), and serves six stakeholder groups: farmers, citizens, disaster management authorities, aviation, marine users, and researchers.

How we built it

  • Frontend: HTML5, CSS3, JavaScript with GSAP for animation, and the Web Speech API for voice input/output.
  • Backend: Python (FastAPI) orchestrating the query pipeline.
  • Language layer: Bhashini API for Automatic Speech Recognition, translation, and text-to-speech across Indian languages.
  • Data sources: Open-Meteo (aggregating NOAA GFS, ECMWF, DWD models), with a roadmap to IMD and ISRO-MOSDAC satellite data and NDMA's Sachet alert system.
  • AI pipeline (the core design choice): we use a dual-model verification architecture. Gemini generates the natural-language response, grounded strictly in retrieved weather data — but every numeric and factual claim is then independently cross-checked by Claude against the source data before it reaches the user. If there's a mismatch, the response is rejected and regenerated rather than shown. This is inspired by Retrieval-Augmented Generation research (Lewis et al., 2020) and hallucination-mitigation literature (Huang et al., 2023) — the goal was to make sure an AI tool giving weather and disaster advice never confidently states something wrong.
  • Routing: a rule-based query classifier splits requests into a fast path (simple lookups, cached/templated responses) and a slow path (reasoning-heavy advisory queries), keeping latency low for most queries.
  • Storage: PostgreSQL for structured data, query logs, and cached forecasts.
  • Deployment: Docker (with Kubernetes planned for scale-out).

Challenges we ran into

  • Avoiding AI hallucination in a high-stakes domain. A wrong weather claim isn't a minor bug — it can mean a farmer sprays pesticide before rain or a fisherman misses a storm warning. Designing the two-stage generate-then-verify pipeline, rather than trusting a single LLM call, was the hardest and most important architectural decision we made.
  • Balancing latency against accuracy. Running retrieval, generation, and independent verification in sequence adds latency. We addressed this with response caching and a fast-path/slow-path split so simple queries don't pay the full pipeline cost.
  • Designing for low-connectivity, low-literacy users, not just app users — which pushed us toward WhatsApp and IVR channels instead of a smartphone-only design.
  • Restricted access to official meteorological data. IMD's API access is permissioned and increasingly restricted. We designed around this by starting with open data (Open-Meteo) and planning phased integration with IMD/NDMA in parallel.
  • Multilingual quality variance, especially for lower-resource Indian languages, which led us to plan a phased language rollout starting with high-resource languages.

What we learned

We learned that in domains like disaster response and agriculture, trustworthiness matters more than fluency — an AI system has to be honest about uncertainty and verifiably correct, not just articulate. We also learned a lot about multilingual NLP infrastructure in the Indian context (Bhashini, IndicTrans2) and about designing systems for users who may have no smartphone, no literacy with forecast bulletins, and no reliable internet — constraints that shaped nearly every technical decision we made.

What's next for MausamGPT

  • Formal integration with IMD and NDMA's Sachet alert system
  • Dedicated advisory modules for aviation (METAR/TAF-based briefings) and marine users
  • WebSocket/MQTT-based real-time push alerts
  • Kubernetes-based scaling to handle demand spikes during extreme weather events, when reliability matters most

Built With

Share this project:

Updates

Submission history