WeatherGPT: Conversational AI for Weather Forecasting, Alerts, and Climate Information

1. Executive Snapshot

Summary

The WeatherGPT project aims to solve the problem of fragmented meteorological information dissemination across India [PS]. Currently, weather data is scattered across multiple portals, bulletins, satellite feeds, and forecast systems, making it difficult for citizens, farmers, researchers, and disaster managers to obtain timely, actionable insights [PS]. WeatherGPT will serve as a multi-channel, voice-enabled conversational AI platform that ingests real-time meteorological datasets, numerical weather prediction (NWP) model outputs (such as GFS/WRF), and disaster warning feeds [PS]. It delivers location-specific weather updates, early disaster warnings, agromet advisories, aviation briefings, and historical climate trend analyses in multiple Indian languages via natural language querying and voice interaction [PS].

Metadata

Fundamental System Objective

To build a scalable, multilingual, and voice-enabled conversational AI platform that synthesizes multi-source meteorological datasets, NWP model outputs, and CAP alerts to provide instant, location-tailored, and actionable weather intelligence to citizens, farmers, disaster response authorities, and domain specialists [INFERENCE].

2. Problem Deconstruction

The problem statement has been broken down into atomic, unambiguous requirements preserving all explicit constraints [PS]:

1. REQ-01: Multi-Source Meteorological Data Ingestion [PS]

   • Source Text: "Backend integration with meteorological databases, website and APIs." [PS] & "Integration with numerical weather prediction (NWP) models such as GFS/WRF." [PS]
   • Interpretation: The system must dynamically pull and ingest observational data, station feeds, satellite/radar products, and NWP model outputs (GFS/WRF) via REST APIs, web scrapers, or database connectors [INFERENCE].
   • Implementation Implication: Requires asynchronous data ingestors with caching layer and spatial query capabilities [INFERENCE].
   • Tag: [PS]

2. REQ-02: Real-Time Weather Information Retrieval & Querying [PS]

   • Source Text: "Real-time weather information retrieval." [PS] & "Natural language querying for weather forecasts." [PS]
   • Interpretation: Users can ask natural language questions (e.g., "Will it rain in Pune today at 4 PM?") and receive real-time, accurate weather metrics [INFERENCE].
   • Implementation Implication: Intent classification engine mapping user entities (location, time, parameter) to structured API queries [INFERENCE].
   • Tag: [PS]

3. REQ-03: Extreme Weather Alerts & Early Warning Dissemination [PS]

   • Source Text: "Extreme weather alerts and early warning dissemination." [PS]
   • Interpretation: Broadcast and push urgent warning alerts (cyclones, heavy rainfall, heatwaves, lightning) formatted according to standards like CAP (Common Alerting Protocol) [INFERENCE].
   • Implementation Implication: Integration with NDMA/IMD CAP alert streams or MQTT event brokers for immediate push notifications [INFERENCE].
   • Tag: [PS]

4. REQ-04: Location-Based Forecasting & Advisory Generation [PS]

   • Source Text: "Location-based forecasting and advisory generation." [PS]
   • Interpretation: Generate hyper-local forecasts using device GPS or geocoded user queries and produce specialized domain advisories (e.g., farming, marine, urban planning) [INFERENCE].
   • Implementation Implication: Reverse geocoding module mapped to IMD grid points or block-level forecast databases [INFERENCE].
   • Tag: [PS]

5. REQ-05: Multilingual Support for Indian Languages [PS]

   • Source Text: "Multilingual support for Indian languages." [PS]
   • Interpretation: Support input and output queries in major Indian scheduled languages (e.g., Hindi, Tamil, Telugu, Marathi, Bengali, Kannada) [INFERENCE].
   • Implementation Implication: Neural machine translation (NMT) models or multilingual LLM prompting with language detection [INFERENCE].
   • Tag: [PS]

6. REQ-06: Climate Trend & Historical Weather Analysis [PS]

   • Source Text: "Climate trend and historical weather analysis." [PS]
   • Interpretation: Answer analytical queries regarding long-term temperature changes, rainfall anomalies, monsoon trends, and historical weather extremes [INFERENCE].
   • Implementation Implication: Historical time-series database (e.g., PostgreSQL/TimescaleDB) and statistical/analytics query generator [INFERENCE].
   • Tag: [PS]

7. REQ-07: Voice-Enabled Interaction for Rural Accessibility [PS]

   • Source Text: "Voice-enabled interaction for rural accessibility." [PS] & "Voice-enabled interaction for rural accessibility" (Evaluation Parameter) [PS]
   • Interpretation: Rural users (e.g., farmers, fishermen) can speak to the application in their local dialect/language and listen to synthesized audio responses [INFERENCE].
   • Implementation Implication: Speech-to-Text (STT) and Text-to-Speech (TTS) pipeline supporting low-bandwidth compressed audio streaming [INFERENCE].
   • Tag: [PS]

8. REQ-08: Scalable Architecture Supporting Protocols (WIS 2.0, MQTT, WebSocket) [PS]

   • Source Text: "MQTT / WIS2.0 / WebSocket" [PS] & "Scalable architecture supporting real-time data ingestion." [PS]
   • Interpretation: Backend must conform to WMO Information System 2.0 (WIS 2.0) standards using MQTT publish/subscribe mechanisms and WebSocket for real-time mobile push updates [INFERENCE].
   • Implementation Implication: Deployment of MQTT broker (Mosquitto/EMQX) subscribing to WIS 2.0 topics and persistent WebSocket server [INFERENCE].
   • Tag: [PS]

9. REQ-09: Mobile-First Conversational Platform [PS]

   • Source Text: "A mobile-based conversational AI platform." [PS]
   • Interpretation: The primary interface must be a responsive mobile application (React Native/Flutter/PWA) capable of offline caching and low-latency interaction [INFERENCE].
   • Implementation Implication: Cross-platform mobile deployment targeting Android (due to rural India penetration) with voice recording and location services [INFERENCE].
   • Tag: [PS]

10. REQ-10: Multi-Domain Decision Support Systems [PS]

    • Source Text: "Intelligent weather decision-support system for agriculture, aviation, marine, and urban planning." [PS]
    • Interpretation: Domain-specific persona filters that translate raw weather metrics into actionable context (e.g., crop spraying windows for farmers, crosswind/visibility for pilots, wave height for fishermen, drainage runoff for smart cities) [INFERENCE].
    • Implementation Implication: Rule engines combined with domain-conditioned RAG (Retrieval-Augmented Generation) templates [INFERENCE].
    • Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Connectivity Realities in Rural/Coastal Areas [KNOWN]

   • Impact: Rural agricultural belts and marine offshore zones frequently suffer from intermittent 2G/3G connectivity or packet loss [KNOWN].
   • Design Response: Implement an offline-first architecture with local SQLite/IndexedDB caching, lightweight JSON payloads, and audio compression (Opus format) [INFERENCE].

2. Language & Literacy Barriers [KNOWN]

   • Impact: Over 50% of rural agricultural workers in India rely on spoken local dialects and cannot easily parse written text or complex meteorological terms [KNOWN].
   • Design Response: Voice-first interaction model with automatically generated audio responses in regional Indian languages (Hindi, Marathi, Telugu, Tamil, etc.) using lightweight TTS [INFERENCE].

3. Hardware Constraints on Field Devices [KNOWN]

   • Impact: Target users often operate entry-level smartphones with limited storage (<32 GB) and constrained RAM (<3 GB) [KNOWN].
   • Design Response: Avoid heavy client-side AI execution. Keep client bundle size under 15 MB and offload speech recognition / LLM inference to a cloud pipeline [INFERENCE].

4. Heterogeneous Meteorological Data Formats [RESEARCHED]

   • Impact: Weather data exists in diverse formats including GRIB2 (GFS/WRF forecasts), NetCDF (historical gridded data), CAP XML (alerts), GeoTIFF (satellite/radar), and JSON/REST APIs (station observations) [RESEARCHED].
   • Design Response: Build a unified Data Abstraction Layer that transforms heterogeneous formats into a standardized spatial-temporal schema in PostgreSQL/PostGIS [INFERENCE].

5. Rate Limits & IP Whitelisting on Government APIs [RESEARCHED]

   • Impact: Official IMD APIs (api.imd.gov.in) require static IP whitelisting and enforce strict rate limits during extreme weather events [RESEARCHED].
   • Design Response: Implement a backend caching proxy (Redis) with exponential backoff and fallback mock data triggers during hackathon demonstrations [INFERENCE].

6. Research Questions

1. How does IMD structure its public REST APIs and CAP alert payloads, and what authentication/whitelisting mechanisms are required? [RESEARCHED]
2. What is the exact topic hierarchy and message format specified by WMO for WIS 2.0 MQTT brokers? [RESEARCHED]
3. How are NWP GFS (0.25° global) and WRF (3 km high-resolution regional) model outputs processed from GRIB2 files into rapid spatial point queries? [RESEARCHED]
4. Which open-source open-weights LLMs (e.g., Llama 3, Sarvam AI / IndicBERT, Gemma) offer the highest accuracy for Indian language translation and domain-specific intent extraction? [RESEARCHED]
5. How does the Meghdoot Agromet DSS format location-specific crop advisories across Indian agricultural districts? [RESEARCHED]

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

1. IMD Mausam App

• Description: Official mobile application launched by IMD providing current weather, 7-day city forecasts, nowcasts, and severe weather warnings [RESEARCHED].
• Strengths: Highly authoritative data source directly backed by MoES observational network [RESEARCHED].
• Weaknesses: Static GUI-based interface, rigid tab navigation, lack of natural language querying, limited conversational AI capability, and clunky voice support [RESEARCHED].
• Relevance: Serves as the primary underlying data authority [INFERENCE].
• Gap: Lacks interactive, natural language voice-driven Q&A and cross-domain intelligent synthesis [INFERENCE].
• Source: IMD Official Release [RESEARCHED]
• Confidence Tag: [RESEARCHED]

2. Meghdoot App (IMD / IITM / ICRISAT)

• Description: Dedicated agromet advisory mobile application providing district/block-level crop-weather advisories across India [RESEARCHED].
• Strengths: Directly addresses farmers with tailored crop management advice in 12 Indian languages [RESEARCHED].
• Weaknesses: Non-conversational text/tabular UI; users must manually select state, district, crop, and date; lacks interactive voice bot capability [RESEARCHED].
• Relevance: Baseline benchmark for agricultural weather advisories [INFERENCE].
• Gap: No conversational AI engine to answer specific follow-up questions like "Can I spray pesticide on my cotton crop today evening if it rains in 3 hours?" [INFERENCE].
• Source: ICRISAT / IMD Meghdoot Documentation [RESEARCHED]
• Confidence Tag: [RESEARCHED]

3. Sachet Portal / NDMA CAP Alert System

• Description: National Disaster Management Authority (NDMA) platform for disseminating geotargeted early warning alerts via SMS and mobile app using CAP standard [RESEARCHED].
• Strengths: Excellent multi-hazard coverage and direct integration with telecom operators [RESEARCHED].
• Weaknesses: One-way alert broadcast; zero conversational capability or context exploration [INFERENCE].
• Relevance: Standard feed source for extreme weather early warnings [INFERENCE].
• Gap: Users cannot query the alert for actionable safety steps or specific localized evacuation guidance [INFERENCE].
• Source: NDMA CAP Portal [RESEARCHED]
• Confidence Tag: [RESEARCHED]

4. Commercial AI Weather Chatbots (e.g., Tomorrow.io AI / ClimaCell)

• Description: Proprietary commercial weather platforms offering AI-assisted weather insights for global enterprise customers [KNOWN].
• Strengths: Sleek user interfaces, high-resolution global models, and strong NLP capabilities [KNOWN].
• Weaknesses: Expensive SaaS licensing, lack of deep integration with official Indian IMD grid systems, zero support for Indic rural dialects or IMD Agromet bulletins [INFERENCE].
• Relevance: Architecture design reference for LLM + Weather data integration [INFERENCE].
• Gap: Not tailored to Indian regional languages, IMD specific standards, or rural Indian low-bandwidth environments [INFERENCE].
• Source: Commercial documentation [KNOWN]
• Confidence Tag: [KNOWN]

9. Candidate Solution Approaches

10. Recommended Approach & Why

The recommended architecture is Approach 3: Hybrid AI Architecture (LLM Intent Extractor & Structured Spatial RAG Engine) [INFERENCE].

Why Alternatives Were Rejected

1. Rejection of Approach 1 (Rule-Based System): While fast, pure rule-based systems break whenever users speak in natural, non-standard Indic dialects or phrase complex queries (e.g., "If it rains tomorrow morning in Nashik, should I harvest my grapes today?"). It fails to satisfy the core SIH mandate for an "intelligent conversational platform" [PS].
2. Rejection of Approach 2 (Pure Unstructured LLM / Web-RAG): LLMs inherently struggle with precise spatial-temporal numeric calculations and are known to hallucinate numerical values (e.g., predicting 45°C instead of 35°C or inventing non-existent rain warnings) [KNOWN]. In disaster management and aviation weather, hallucinated numbers can lead to catastrophic consequences [INFERENCE].

Why Approach 3 Wins

• Guaranteed Numeric Factual Accuracy: The LLM never generates weather data from its parametric memory. Instead, it converts user text/speech into a structured JSON query payload (Location: "Pune", Time: "Next 24 Hours", Domain: "Agriculture", Parameter: "Precipitation") [INFERENCE].
• Sub-Second Latency: Structured database lookups against pre-ingested GFS/IMD caches take <50ms. The LLM is only used for lightweight intent extraction and final natural language voice synthesis [INFERENCE].
• Safety Guardrails: If an extreme alert (CAP XML) exists for the user's location, the system overrides normal responses and injects high-priority safety instructions [INFERENCE].

11. Technical Architecture

+---------------------------------------------------------------------------------------------------+
|                                  USER INTERFACE LAYER (MOBILE & PWA)                              |
|   React Native / Flutter App  |  Web Speech API / Opus Audio Stream  |  Geolocation SDK           |
+---------------------------------------------------------------------------------------------------+
                                           |
                                  (HTTPS / WebSocket)
                                           v
+---------------------------------------------------------------------------------------------------+
|                                     API GATEWAY & ROUTER LAYER                                    |
|   FastAPI Gateway  |  JWT / OAuth2 Auth  |  Rate Limiter  |  Locale & Dialect Detector            |
+---------------------------------------------------------------------------------------------------+
                                           |
                    +----------------------+----------------------+
                    |                                             |
                    v                                             v
+---------------------------------------+   +-------------------------------------------------------+
|  SPEECH & TRANSLATION ENGINE          |   |  INTENT & ENTITY EXTRACTION ENGINE (LLM)             |
|  - Whisper / Bhashini / IndicASR      |   |  - Llama-3-8B / Ollama / OpenAI GPT-4o-Mini          |
|  - Speech-to-Text (STT)               |   |  - Function Calling / JSON Output Enforcement        |
|  - Text-to-Speech (TTS)               |   |  - Extracts: {location, time_range, domain, metric}  |
+---------------------------------------+   +-------------------------------------------------------+
                    |                                             |
                    +----------------------+----------------------+
                                           |
                                           v
+---------------------------------------------------------------------------------------------------+
|                                 SPATIAL DATA & RAG ORCHESTRATOR                                   |
|   - Geocoding Resolver (Nominatim / PostGIS Reverse Geocode)                                      |
|   - Spatial Point-in-Polygon Query (Mapped to IMD District / Grid Point)                          |
|   - Domain Persona Injector (Agromet / Aviation METAR / Marine / Urban Storm)                     |
+---------------------------------------------------------------------------------------------------+
                                           |
        +----------------------------------+----------------------------------+
        |                                  |                                  |
        v                                  v                                  v
+-----------------------+     +-----------------------+     +-----------------------------------+
| IMD / GFS CACHE DB    |     | CAP ALERT BROKER      |     | HISTORICAL CLIMATE DB             |
| - PostgreSQL/PostGIS  |     | - MQTT (Mosquitto)    |     | - TimescaleDB / NetCDF Engine     |
| - Redis (Realtime)    |     | - WIS 2.0 Topic Sub   |     | - Decadal Anomalies               |
+-----------------------+     +-----------------------+     +-----------------------------------+
        ^                                  ^                                  ^
        |                                  |                                  |
+---------------------------------------------------------------------------------------------------+
|                                DATA INGESTION & PIPELINE LAYER                                    |
|   - IMD REST Ingestor  |  WIS 2.0 MQTT Subscriber  |  NOAA GFS GRIB2 Parser  |  CAP XML Ingestor    |
+---------------------------------------------------------------------------------------------------+

Detailed Data Flow Narrative

1. User Request: A farmer in Nashik presses the voice button on the mobile app and speaks in Marathi: "आज संध्याकाळी पाऊस पडेल का? मला पिकावर औषध फवारायचे आहे." ("Will it rain this evening? I need to spray pesticide on my crop.") [INFERENCE].
2. Audio Ingestion & STT: The compressed audio (Opus format over WebSocket) reaches the FastAPI backend and is converted to text using an Indic Speech-to-Text engine (Bhashini API or OpenAI Whisper) [INFERENCE].
3. Intent & Entity Extraction: The transcribed text is sent to the LLM Query Engine. Using strict JSON function calling, the LLM extracts:
   • Intent: WEATHER_FORECAST_AND_AGRO_ADVISORY
   • Location: Nashik, Maharashtra
   • Time_Range: Today Evening (16:00 - 20:00 IST)
   • Domain: Agriculture (Sub-category: Pesticide Spraying) [INFERENCE].
4. Spatial Resolution & Querying: The orchestrator converts "Nashik" into coordinates (lat/long) via PostGIS, fetches real-time hourly GFS/IMD forecast data from Redis/PostgreSQL, and checks the CAP Alert MQTT broker for active severe weather warnings [INFERENCE].
5. Domain Context Synthesis: The database returns: Rain Probability: 85%, Expected Rainfall: 12mm between 17:00-19:00 IST, Wind Speed: 18 km/h. The Agromet Rule Engine evaluates: Pesticide spraying not recommended when rain probability > 40% or wind > 15 km/h [INFERENCE].
6. Response Generation & TTS: The synthesized context + raw metrics are passed back to the LLM to format a concise, empathetic response in Marathi. The text is passed to an Indic TTS model, streaming audio back to the mobile app [INFERENCE].

12. Technology Stack

13. AI/ML Justification

AI Component 1: Intent & Entity Extractor (LLM)

• Purpose: Extract structured parameters (location, temporal range, weather metric, user persona) from raw unstructured user queries in natural language [INFERENCE].
• Model Type: Fine-tuned Llama-3-8B-Instruct or GPT-4o-Mini with JSON schema mode / Function Calling [INFERENCE].
• Input: Unstructured text query (e.g., "Will it hail in Shimla over the weekend?") [INFERENCE].
• Output: JSON payload {"location": "Shimla", "start_date": "2026-09-05", "end_date": "2026-09-06", "phenomenon": "hail"} [INFERENCE].
• Data Required: 500+ annotated multi-domain weather queries in English and Indian languages [ASSUMPTION].
• Why AI is Preferable: Simple regex or rule parsers fail on complex, dialectal, or multi-intent queries [INFERENCE].
• Evaluation Metric: Entity extraction F1-score (>0.92 target) [INFERENCE].
• Failure Mode: Unrecognized location -> Prompt user for clarification or fallback to device GPS [INFERENCE].
• Human Oversight: High-risk disaster parameters are cross-checked against CAP alert tables before generating response [INFERENCE].

AI Component 2: Indic Speech Processing (STT / TTS)

• Purpose: Convert spoken audio queries into text and synthesize written responses back into clear audio for rural users [PS].
• Model Type: Whisper-Medium fine-tuned on Indic speech / AI4Bharat Bhashini Models [RESEARCHED].
• Input: Audio stream (Opus/WAV) from mobile device [INFERENCE].
• Output: Transcribed Indic text / Audio output byte stream [INFERENCE].
• Data Required: Pre-trained Indic speech corpora [RESEARCHED].
• Why AI is Preferable: Deterministic algorithms cannot process acoustic human speech variations [KNOWN].
• Evaluation Metric: Word Error Rate (WER < 12%) [INFERENCE].
• Failure Mode: Noisy audio background -> Return text transcription to screen and request voice repeat [INFERENCE].
• Human Oversight: On-screen transcript allows user to visually inspect speech recognition accuracy [INFERENCE].

Explicit Non-AI Component: Numeric Forecast Calculation

• Decision: DO NOT USE AI / LLMs TO CALCULATE OR PREDICT NUMERIC FORECAST METRICS DIRECTLY [INFERENCE].
• Justification: Weather prediction requires solving complex physical thermodynamic differential equations (run by supercomputers executing WRF/GFS models) [KNOWN]. LLMs predicting future temperature or precipitation values produce dangerous hallucinations [INFERENCE]. All numerical weather metrics MUST come directly from IMD/NOAA database tables [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Database Schema (Simplified SQL Definition)

-- Weather Station Current Observations
CREATE TABLE station_observations (
    station_id VARCHAR(50) PRIMARY KEY,
    station_name VARCHAR(100),
    geom GEOMETRY(Point, 4326),
    temp_c NUMERIC(4,1),
    humidity_pct INT,
    wind_speed_kmh NUMERIC(4,1),
    rainfall_24h_mm NUMERIC(5,1),
    recorded_at TIMESTAMPTZ NOT NULL
);

-- NWP Grid Forecasts (GFS / WRF Ingest)
CREATE TABLE nwp_grid_forecasts (
    id BIGSERIAL PRIMARY KEY,
    model_name VARCHAR(20), -- 'GFS' or 'WRF'
    grid_point GEOMETRY(Point, 4326),
    forecast_time TIMESTAMPTZ NOT NULL,
    temp_2m NUMERIC(4,1),
    precip_prob_pct INT,
    precip_amount_mm NUMERIC(5,1),
    wind_u NUMERIC(4,1),
    wind_v NUMERIC(4,1),
    created_at TIMESTAMPTZ DEFAULT NOW()
);
CREATE INDEX idx_nwp_spatial ON nwp_grid_forecasts USING GIST(grid_point);
CREATE INDEX idx_nwp_time ON nwp_grid_forecasts(forecast_time);

-- CAP Disaster Alerts
CREATE TABLE cap_alerts (
    alert_id VARCHAR(100) PRIMARY KEY,
    headline TEXT,
    severity VARCHAR(20), -- 'Extreme', 'Severe', 'Moderate'
    affected_zone GEOMETRY(Polygon, 4326),
    instruction TEXT,
    effective_from TIMESTAMPTZ,
    expires_at TIMESTAMPTZ
);

Offline & Low-Network Optimization Strategy

1. Local SQLite / PWA Cache: The mobile client caches the 7-day forecast for the user's saved location locally [INFERENCE].
2. Bandwidth Reduction: Voice streams are compressed using the Opus codec at 16 kbps (reducing data usage by 80% compared to raw PCM WAV) [INFERENCE].
3. Payload Compression: All backend REST and WebSocket JSON responses are gzip/brotli compressed [INFERENCE].
4. Offline Mode UX: When network connection drops below 2G speeds:
   • The app notifies the user: "Working Offline. Showing cached forecast from 2 hours ago." [INFERENCE].
   • Critical safety rules (e.g., standard lightning safety rules) remain locally available on device storage [INFERENCE].

15. Security, Privacy & Governance

1. Personal & Location Data Protection [INFERENCE]

   • Precise GPS coordinates are rounded to 2 decimal places (~1.1 km resolution) before storing in backend logs to prevent user tracking [INFERENCE].
   • Compliance with India's Digital Personal Data Protection (DPDP) Act 2023 [KNOWN].

2. Authentication & Authorization [INFERENCE]

   • Citizen interfaces use anonymous JWT tokens tied to device installation IDs [INFERENCE].
   • Administrative dashboards (for disaster managers and IMD operators) require multi-factor authentication (MFA) via OAuth2/Keycloak [INFERENCE].

3. Encryption Standards [KNOWN]

   • Data in Transit: TLS 1.3 enforced across all REST, WebSocket, and MQTT protocols [KNOWN].
   • Data at Rest: AES-256 encryption on PostgreSQL database volumes and Redis instances [KNOWN].

4. Liability & AI Failure Disclaimers [INFERENCE]

   • All AI-generated advisory outputs contain an explicit disclaimer: "WeatherGPT is an automated decision support prototype based on official IMD/MoES guidance. For official emergency evacuations, consult local district authorities." [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD & DEMO)

• Core Conversational Engine: React Native mobile web application with text and voice query input [PS].
• Multilingual Support: Dual language support (Hindi + English) with automatic query translation [PS].
• Real-Time Data Ingestion: Simulated/Live API integration for city weather forecasts and 7-day guidance [PS].
• WIS 2.0 / MQTT Simulation Engine: Working MQTT publisher streaming CAP extreme weather alerts to mobile clients in real time [PS].
• Domain Personas: Agriculture (crop advisory) and Aviation (METAR query) persona filters [PS].
• Historical Climate Explorer: Visual chart rendering decadal temperature/rainfall trends for selected Indian cities [PS].

Described Only / Out of Scope (PRODUCTION FUTURE)

• Integration with proprietary live Doppler Radar / Satellite raw binary feeds [INFERENCE].
• Support for all 22 scheduled Indian regional languages [INFERENCE].
• Direct WhatsApp / Telecom SMS broadcast integration [INFERENCE].
• Automated drone / IoT weather station hardware pairing [INFERENCE].

17. Live Demo Simulation Plan

During the SIH live judging, cloud connectivity or live government API endpoints may be restricted or rate-limited [INFERENCE]. The team will use a reliable Live Data Simulation Architecture:

1. Simulated WIS 2.0 / MQTT Stream:

   • Run a local Mosquitto MQTT broker on the demonstration laptop [INFERENCE].
   • Use a Python script to inject a simulated Severe Weather CAP Alert (e.g., "Extreme Rain & Cloudburst Warning for District Pune") into topic wis2/india/imd/alerts/extreme [INFERENCE].
   • Demonstrate the mobile app instantly receiving a high-priority banner pop-up and voice warning over WebSocket [INFERENCE].

2. Simulated NWP Model Ingestion:

   • Pre-pack actual historical NOAA GFS GRIB2 data files for a sample monsoon period [INFERENCE].
   • Show the system parsing GRIB2 grid points and serving sub-second spatial queries without relying on external internet latency during judging [INFERENCE].

3. Transparency Principle:

   • Clearly label the test control panel: "SIH Live Demonstration Environment - Streaming Real-Time Mock Ingestion Feeds" [INFERENCE].

18. Implementation Roadmap

Project Timeline: Target Deadline - 30 September 2026 [PS]

+------------------------------------------------------------------------------------+
| PHASE 1: DATA ARCHITECTURE & INGESTION (Weeks 1-2)                                  |
| - Setup PostgreSQL/PostGIS & Redis schema.                                        |
| - Build API scrapers/connectors for Open-Meteo, IMD APIs, and GFS open data.      |
| - Deploy local Mosquitto MQTT broker for WIS 2.0 testing.                          |
+------------------------------------------------------------------------------------+
                                         |
                                         v
+------------------------------------------------------------------------------------+
| PHASE 2: AI INTENT & DOMAIN ENGINE (Weeks 3-4)                                     |
| - Implement FastAPI server with Groq/Ollama Llama-3 integration.                  |
| - Develop JSON function calling schemas for weather intent extraction.            |
| - Build Agromet, Aviation, and CAP alert prompt templates.                        |
+------------------------------------------------------------------------------------+
                                         |
                                         v
+------------------------------------------------------------------------------------+
| PHASE 3: VOICE & MULTILINGUAL PIPELINE (Weeks 5-6)                                 |
| - Integrate Bhashini / Whisper STT and Indic TTS models.                           |
| - Test Marathi/Hindi audio query round-trip latency (<2 seconds).                  |
| - Optimize Opus audio streaming over WebSocket.                                    |
+------------------------------------------------------------------------------------+
                                         |
                                         v
+------------------------------------------------------------------------------------+
| PHASE 4: MOBILE APP & DEMO POLISH (Weeks 7-8)                                      |
| - Complete React Native UI with dynamic chart components (Recharts/Victory).      |
| - Integrate map layers (Leaflet/Mapbox) for weather overlays.                       |
| - End-to-end testing, demo script rehearsing, and Docker packaging.                |
+------------------------------------------------------------------------------------+

19. Testing & Validation Plan

1. Functional Testing:

   • Test: Query 50 distinct cities across India for current weather and 7-day forecast.
   • Pass Criteria: 100% correct match between database query results and displayed UI metrics [INFERENCE].

2. AI Intent & False Positive Testing:

   • Test: Supply 100 non-weather queries (e.g., "Who won the cricket match?") and ambiguous queries (e.g., "Is it hot in Delhi?").
   • Pass Criteria: Non-weather queries gracefully rejected; weather queries parsed with >90% intent accuracy [INFERENCE].

3. False Negative Alert Testing:

   • Test: Inject critical CAP alerts into the MQTT broker while the user is actively asking about normal weather.
   • Pass Criteria: Active disaster warning MUST interrupt standard chat and display emergency banner [INFERENCE].

4. Performance & Latency Testing:

   • Test: Benchmark text query response time under simulated 3G network conditions (300ms latency).
   • Pass Criteria: End-to-end text query response < 1.5 seconds; voice query response < 3.0 seconds [INFERENCE].

5. Offline Resiliency Testing:

   • Test: Turn off device Wi-Fi/Cellular data and issue a query.
   • Pass Criteria: App displays cached forecast data with explicit timestamp indicator without crashing [INFERENCE].

20. Risk Register

21. Cost & Scalability

Hackathon MVP Build

• Cloud Hosting: Free tier Vercel/Render (Backend) + Free tier Supabase (PostgreSQL/PostGIS) = $0 [ASSUMPTION].
• LLM API: Groq API developer free tier (30 requests/min) = $0 [ASSUMPTION].
• Voice APIs: AI4Bharat Bhashini Open API / Local Whisper = $0 [RESEARCHED].
• Total Hackathon Cost: ₹0 (Free Tier Stack) [ASSUMPTION].

Production System (National Deployment - MoES Scale)

• Infrastructure: Distributed Kubernetes Cluster (AWS/NIC Cloud) with 4x NVIDIA T4 GPUs for STT/TTS = ~$2,500 / month [ASSUMPTION].
• High Availability DB: Managed PostgreSQL Cluster with TimescaleDB read-replicas = ~$800 / month [ASSUMPTION].
• MQTT Broker: Enterprise EMQX Cluster supporting 10,000,000 simultaneous mobile connections = ~$1,200 / month [ASSUMPTION].
• Scalability Target: Capable of handling 50,000 requests per second during active cyclone/monsoon alert windows using CDN edge caching [INFERENCE].

22. Differentiation

23. Future Scope

1. Edge AI Deployment: Quantize STT and intent-parsing models to run directly on-device (Android NPU) for 100% offline voice operation in remote disaster zones [INFERENCE].
2. Crowdsourced Weather Observations: Allow farmers and citizens to upload geotagged photo reports of hail, localized flooding, or crop damage to validate NWP models [INFERENCE].
3. Hyper-Local IoT Micro-Weather Station Mesh: Connect low-cost LoRaWAN field sensors directly into the WeatherGPT MQTT ingestor for sub-kilometer microclimate tracking [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions & Items Needing Validation

• [ASSUMPTION] The hackathon judging panel accepts pre-processed/simulated NOAA GFS data in place of whitelisted live IMD production API tokens if external internet is restricted.
• [ASSUMPTION] Groq Cloud API free tier rate limits will accommodate live continuous query testing during presentation slots.
• [NEEDS VALIDATION] Official IP whitelisting approval timelines from IMD (api.imd.gov.in) for student hackathon teams.
• [NEEDS VALIDATION] Operational status and public endpoint accessibility of official WMO WIS 2.0 Global Broker MQTT topics within India.

Final Recommendation

Build WeatherGPT as a Hybrid AI Platform featuring a React Native PWA Mobile Frontend, FastAPI Orchestrator, PostgreSQL/PostGIS Spatial Database, Mosquitto MQTT Alert Ingestor, and Bhashini/Whisper Indic Voice Engine. Keep numeric weather processing 100% deterministic while leveraging LLMs exclusively for intent parsing and natural language response synthesis.

Recommended Stack

• Frontend: React Native (Expo) + Web Speech API + Leaflet Maps
• Backend: Python 3.12 + FastAPI + Uvicorn
• Databases: PostgreSQL 16 + PostGIS + TimescaleDB + Redis
• Messaging / Ingestion: Eclipse Mosquitto (MQTT) + WebSockets
• AI / Speech: Ollama / Groq (Llama-3-8B) + AI4Bharat Bhashini STT/TTS

Biggest Dependency

Access to structured, location-indexed weather forecast feeds (NOAA GFS / Open-Meteo fallback when official IMD APIs are rate-limited or require static IP whitelisting) [RESEARCHED].

Biggest Risk

Speech recognition failure or high audio translation latency when rural users speak in low-bandwidth environments with heavy background noise [INFERENCE].

Biggest Technical Challenge

Processing heterogeneous meteorological data formats (GRIB2 binary grids, NetCDF historical files, CAP XML alerts) into a unified sub-50ms spatial PostGIS lookup engine [INFERENCE].

First 5 Actions

1. Action 1: Initialize Git repository and spin up PostgreSQL/PostGIS + Redis container stack using Docker Compose.
2. Action 2: Develop Python ingestion script to fetch and parse GFS/Open-Meteo weather JSON and store grid points in PostGIS.
3. Action 3: Implement FastAPI endpoint integrating Llama-3 (via Groq/Ollama) with strict Pydantic JSON schema function calling for intent extraction.
4. Action 4: Integrate Bhashini / Whisper Speech-to-Text and Indic TTS pipeline with WebSocket audio streaming.
5. Action 5: Build React Native voice-enabled chat interface with real-time alert popups and historical climate charts.

Validation Before Coding

Verify that the spatial query execution time on PostGIS for point-in-polygon district matching takes under 30ms on developer hardware.

Buildability Score /100

88 / 100 (Highly feasible for a skilled 4-6 student team in a hackathon build window using open-source tools, Python ecosystem, and Docker packaging).

Confidence

High (Supported by detailed research of IMD APIs, WIS 2.0 standards, and proven spatial RAG AI architecture patterns).

Research Sources

1. IMD API Management Portal

   • URL: https://api.imd.gov.in
   • Source Type: Official Government Documentation
   • Supported Claim: Confirms official IMD REST API endpoints, station data structures, and whitelisting requirements.
   • Reliability: High (Official MoES/IMD Portal).

2. WMO Information System 2.0 (WIS 2.0) Specification

   • URL: https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-information-system-wis/wis2-overview
   • Source Type: International Standards Organization
   • Supported Claim: Confirms MQTT publish/subscribe architecture, topic hierarchies, and HTTP download patterns for WIS 2.0.
   • Reliability: High (World Meteorological Organization).

3. Meghdoot Agro-Met DSS Architecture Study

   • URL: https://oar.icrisat.org/11917/1/WP370.pdf
   • Source Type: Academic / Research Publication (ICRISAT / IMD)
   • Supported Claim: Details the internal data flow, APIs, and district-level advisory distribution for Indian agriculture.
   • Reliability: High (Peer-reviewed research report).

4. Parliament Question on IMD Weather Forecasting & Mobile Apps

   • URL: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2288239
   • Source Type: Official Press Information Bureau (Government of India)
   • Supported Claim: Confirms operational deployment of Mausam, Meghdoot, Damini, and CAP disaster warning systems.
   • Reliability: High (Official Government Release).