ORCA Marine EcOsystem Reasoning with Collaborative Agents

1. Executive Snapshot

The ORCA (Marine EcOsystem Reasoning with Collaborative Agents) project addresses the challenge of transforming complex, high-volume Earth Observation (EO) satellite products and oceanographic forecasts into actionable, context-aware decision support for coastal and maritime stakeholders [PS]. While agencies like ISRO and ESSO-INCOIS generate vast daily streams of Sea Surface Temperature (SST), Chlorophyll-a, ocean currents, and weather alerts, end users—particularly small-scale traditional fishermen and coastal operators—struggle with multi-portal fragmentation, technical data formats, and lack of localized natural language reasoning [INFERENCE]. ORCA builds an Agentic AI platform that orchestrates specialized AI agents to autonomously interpret multi-lingual user queries, discover and correlate multi-source geospatial datasets, evaluate maritime safety and geofencing risks, and deliver explainable advisories via conversational interfaces, interactive maps, and localized audio [PS].

Fundamental System Objective: To provide an agentic multi-agent conversational decision-support platform that autonomously ingests, correlates, and reasons over satellite Earth observation data, weather forecasts, and oceanographic layers to deliver localized, multi-lingual, explainable advisories and safety notifications for Indian marine stakeholders [PS].

2. Problem Deconstruction

The problem description provided by ISRO explicitly defines core system capabilities and typical user interaction scenarios. Below is the atomic extraction preserving all explicit problem statement requirements [PS] alongside derived technical implications [INFERENCE].

Core Functional & Architectural Requirements

1. Natural Language Query Understanding & Multilingual Support [PS]

   • Requirement: Understand user intent expressed in natural language and automatically identify the language of the query to respond in the same language, with specific emphasis on Indian regional coastal languages [PS].
   • Interpretation: The platform must process text and voice inputs across languages like Tamil, Telugu, Malayalam, Kannada, Marathi, Gujarati, Odia, Bengali, Hindi, and English [INFERENCE].
   • Implementation Implication: Ingest pipeline requires multi-lingual Large Language Model (LLM) embeddings, Automatic Speech Recognition (ASR) tuned for coastal dialects (e.g., Whisper/AI4Bharat Indicator models), and automated language routing [INFERENCE]. Confidence Tag: [PS]

2. Contextual, Multi-Turn Conversation Management [PS]

   • Requirement: Support contextual, multi-turn conversations enabling users to refine queries, explore alternative scenarios, and ask follow-up questions [PS].
   • Interpretation: System must maintain session state, query context, spatial bounds, and user profile parameters (e.g., vessel type, home port) across conversational turns [INFERENCE].
   • Implementation Implication: Stateful dialogue orchestration engine linked to vector session memory and Redis cache [INFERENCE]. Confidence Tag: [PS]

3. Autonomous Multi-Agent Task Decomposition & Orchestration [PS]

   • Requirement: Autonomously interpret user intent, decompose complex requests into executable tasks, coordinate multiple specialized AI agents, and execute tool calling [PS].
   • Interpretation: A single monolithic prompt is insufficient. A lead Orchestrator Agent must break queries down into sub-goals executed by specialized domain agents [INFERENCE].
   • Implementation Implication: Agent framework (e.g., LangGraph / AutoGen) with modular agents: Data Discovery, Ocean Analytics, Weather Intelligence, Geospatial Reasoning, Risk Assessment, and Report Synthesizer [INFERENCE]. Confidence Tag: [PS]

4. Multi-Source Data Retrieval & Correlation [PS]

   • Requirement: Autonomously discover, retrieve, and integrate satellite Earth Observation products (SST, Chlorophyll), GIS layers, weather services, oceanographic observations, and public domain advisories [PS].
   • Interpretation: Platform must interface with heterogeneous open data protocols (ERDDAP, OGC WMS/WFS, NetCDF, GeoTIFF, REST APIs) [INFERENCE].
   • Implementation Implication: Data Abstraction Layer supporting automated spatial-temporal cropping, resampling, and grid alignment [INFERENCE]. Confidence Tag: [PS]

5. Spatial, Temporal, and Contextual Reasoning [PS]

   • Requirement: Perform spatial, temporal, and contextual reasoning by correlating observations from multiple heterogeneous data sources [PS].
   • Interpretation: Rather than returning raw overlays, the system must correlate parameters (e.g., identifying PFZ by overlapping SST thermal fronts with Chlorophyll gradients) [INFERENCE].
   • Implementation Implication: GIS analytics pipeline with Python raster processing libraries (xarray, Rasterio, Shapely) invoked programmatically by AI agents [INFERENCE]. Confidence Tag: [PS]

6. Explainable & Evidence-Based Recommendations [PS]

   • Requirement: Synthesize actionable recommendations through conversational responses, maps, charts, interactive geospatial visualizations, and marine advisories, while explicitly explaining the underlying evidence [PS].
   • Interpretation: Users must receive transparency on why a location is recommended or flagged as dangerous [INFERENCE].
   • Implementation Implication: CoT (Chain-of-Thought) trace capture, bounding box / contour highlight rendering on Leaflet/MapLibre, and explicit citation of data sources (e.g., "EOS-6 OCM-3 pass at 06:00 UTC") [INFERENCE]. Confidence Tag: [PS]

7. Proactive Maritime Safety & Weather Alerts [PS]

   • Requirement: Enhance fishermen safety through proactive alerts for adverse weather, high waves, lightning, cyclones, and hazardous marine conditions [PS].
   • Interpretation: PUSH notification system triggered by threshold monitoring and forecast analysis [INFERENCE].
   • Implementation Implication: Background cron workers evaluating IMD and INCOIS feeds against active user geofences, broadcasting WebPush/SMS alerts [INFERENCE]. Confidence Tag: [PS]

8. Geofencing & Boundary Warnings [PS]

   • Requirement: Provide geofencing-based notifications when approaching International Maritime Boundary Lines (IMBL), restricted waters, Marine Protected Areas (MPA), or ecologically sensitive zones [PS].
   • Interpretation: Prevent accidental international border crossings (e.g., India-Sri Lanka, India-Pakistan maritime boundaries) and illegal fishing in protected zones [INFERENCE].
   • Implementation Implication: Real-time point-in-polygon and distance-to-polyline geospatial spatial queries using PostGIS [INFERENCE]. Confidence Tag: [PS]

9. Route Optimization & Safe Navigation Assistance [PS]

   • Requirement: Assist with route optimization, safe navigation, and operational planning based on prevailing and forecast marine conditions [PS].
   • Interpretation: Calculate optimal vessel routes that minimize fuel burn while avoiding high wave conditions, storms, or restricted zones [INFERENCE].
   • Implementation Implication: A-star or Fast Marching Method (FMM) pathfinding algorithm on dynamic marine velocity/wave cost surface [INFERENCE]. Confidence Tag: [PS]

Atomic User Query Deconstruction

The explicit problem statement outlines 8 typical user query patterns [PS]:

• Query a: "Where is the nearest Potential Fishing Zone today?"
  • Target: Find closest high-probability fish aggregation zone based on SST thermal fronts and chlorophyll concentration relative to current user coordinates [INFERENCE]. Confidence Tag: [PS]
• Query b: "Is it safe to venture into the sea tomorrow morning?"
  • Target: Evaluate short-term ocean state forecast (wave height, swell, wind speed, lightning/cyclone warnings) for next 12-24 hours [INFERENCE]. Confidence Tag: [PS]
• Query c: "What are the tide, weather, and sea conditions near my fishing location?"
  • Target: Contextual point retrieval of tidal height, current velocity, wind vector, and weather nowcast [INFERENCE]. Confidence Tag: [PS]
• Query d: "Are there any lightning or cyclone alerts in my area?"
  • Target: Spatial intersection of user position/buffer with active IMD lightning strike feeds and cyclone warning cones [INFERENCE]. Confidence Tag: [PS]
• Query e: "Which regions show high chlorophyll concentration and favourable sea surface temperature?"
  • Target: Multi-criteria raster filtering (e.g., Chlorophyll > 0.5 mg/m³ AND SST between 26°C-29°C) across selected coastal grid [INFERENCE]. Confidence Tag: [PS]
• Query f: "What is the safest route for a fishing vessel considering weather and sea-state conditions?"
  • Target: Dynamic route generation from Port A to Zone B avoiding weather hazards and high wave zones [INFERENCE]. Confidence Tag: [PS]
• Query g: "Why has fish productivity declined in a particular coastal region?"
  • Target: Multi-temporal trend analysis comparing historical vs current SST anomalies, marine heatwaves, coastal turbidity, or upwelling conditions [INFERENCE]. Confidence Tag: [PS]
• Query h: "Which fishing zones should be avoided due to hazardous marine conditions or geofencing restrictions?"
  • Target: Spatial overlay of hazardous wave/storm areas + IMBL boundaries + Marine Protected Areas [INFERENCE]. Confidence Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. At-Sea Network Connectivity Realities:

   • Onshore / Nearshore (0–12 Nautical Miles): Moderate to low 4G/5G cellular connectivity [RESEARCHED].
   • Offshore (12–50+ Nautical Miles): Complete cellular dead zone [KNOWN]. High-bandwidth satellite internet is cost-prohibitive for small vessel fishermen [RESEARCHED].
   • Operational Constraint: The platform must support a Progressive Web App (PWA) architecture capable of pre-fetching daily PFZ, weather, and geofence vector data while in port, storing it locally in IndexedDB / SQLite WASM for offline display at sea [INFERENCE].

2. Hardware & Power Constraints on Small Vessels:

   • Small motorized boats lack AC power outlets; reliance on low-end Android smartphones charged via 12V boat batteries [RESEARCHED].
   • High sunlight glare on open deck screens; high water exposure [KNOWN].
   • Operational Constraint: Mobile interface must feature high-contrast outdoor UI, large touch targets, voice audio playback, and lightweight raster map tile compression [INFERENCE].

3. Digital & Technical Literacy:

   • Many traditional fishermen cannot interpret technical terms like "2.5m Significant Wave Height" or "0.8 mg/m³ Chlorophyll-a" [RESEARCHED].
   • Operational Constraint: Agent synthesizer must translate numerical metrics into simplified safety levels (Green/Yellow/Red) and plain spoken advisories in local dialects [INFERENCE].

4. Regional Geopolitical & Legal Boundaries:

   • Navigating near the International Maritime Boundary Line (IMBL) in the Palk Strait, Gulf of Mannar, or Sir Creek carries severe detention risks by foreign navies [KNOWN].
   • Operational Constraint: Geofencing boundary alerts must operate with high-precision GPS triggers offline, independent of server connectivity [INFERENCE].

6. Research Questions

1. What official open endpoints or data structures are made public by ISRO (MOSDAC / Bhoonidhi) and ESSO-INCOIS for oceanographic data?
2. How frequently are EOS-6 (Oceansat-3) OCM-3 Chlorophyll and SSTM-1 SST products updated, and what is their spatial resolution?
3. What open geospatial datasets accurately map the Indian International Maritime Boundary Line (IMBL) and Marine Protected Areas (MPAs)?
4. How can an AI agent framework execute spatial raster analysis programmatically without generating hallucinations or invalid syntax?
5. What lightweight routing algorithms effectively handle dynamic ocean surface cost rasters (waves, currents, wind)?

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

1. SAMUDRA Mobile Application (ESSO-INCOIS)

• Description: Official mobile application developed by ESSO-INCOIS (Ministry of Earth Sciences) providing ocean state forecasts, PFZ maps, tsunami warnings, and high wave alerts [RESEARCHED].
• Strengths: Direct access to official INCOIS ocean models, multi-lingual interface, covers PFZ, wave height, and swell alerts [RESEARCHED].
• Weaknesses: Traditional menu-driven application; no natural language interaction; cannot ask complex multi-variable scenarios (e.g., "Why did yield drop in Zone X?"); no route optimization [INFERENCE].
• Relevance: Direct benchmark for ocean data accuracy [RESEARCHED].
• Gap: Lacks Agentic AI reasoning, natural language conversational capabilities, and multi-source synthesis [INFERENCE]. Confidence Tag: [RESEARCHED]

2. ISRO MOSDAC & Bhuvan Geoportals

• Description: ISRO's primary web portals for meteorology/oceanography data (MOSDAC) and national geospatial visualization (Bhuvan) [RESEARCHED].
• Strengths: Authoritative raw satellite products (EOS-6, INSAT-3DS), Web Map Services (WMS/WMTS), OGC compliance [RESEARCHED].
• Weaknesses: Complex GIS interfaces designed for researchers and scientists; unusable on low-end mobile devices in field conditions by non-technical users [INFERENCE].
• Relevance: Primary satellite data provider [RESEARCHED].
• Gap: No conversational AI layer, no automated multi-agent task execution, no simplified field advisory output [INFERENCE]. Confidence Tag: [RESEARCHED]

3. Fisher Friend Mobile Application (MSSRF / Qualcomm)

• Description: Mobile decision-support tool for small-scale fishermen developed by M.S. Swaminathan Research Foundation, providing PFZ, weather, and boundary alerts [RESEARCHED].
• Strengths: Tailored for local coastal fishing communities, multi-lingual audio features [RESEARCHED].
• Weaknesses: Static menu structure; hardcoded rule sets; lacks dynamic route optimization and deep spatial reasoning over raw satellite rasters [INFERENCE].
• Relevance: Benchmark for field usability and local language voice interaction [RESEARCHED].
• Gap: No Agentic multi-agent architecture or real-time spatial correlation capabilities [INFERENCE]. Confidence Tag: [RESEARCHED]

9. Candidate Solution Approaches

Approach 1: Monolithic LLM with RAG and Tool Calling (Single Agent)

• Mechanism: A single LLM (e.g., GPT-4o or Llama-3-70B) equipped with vector search over text advisories and python function tools to query APIs.
• Problem Fit: Low to Moderate. Single LLM struggles with complex multi-step workflows (data retrieval -> raster processing -> spatial join -> safety check -> multi-lingual translation). High risk of context overflow and tool calling hallucinations [INFERENCE].
• Accuracy Potential: Moderate (susceptible to hallucinations during complex math or spatial joins).
• Complexity: Low initial development effort.
• Explainability: Low (reasoning trace is a black box inside one prompt).
• Hackathon Feasibility: High.
• Production Feasibility: Low (expensive API costs, slow latency, unreliable function calling loops). Confidence Tag: [INFERENCE]

Approach 2: Rule-Based GIS Backend with Conversational Intent Router

• Mechanism: Deterministic SQL/PostGIS and Python scripts executed via a hardcoded intent classification bot (e.g., Rasa or Dialogflow).
• Problem Fit: Low. Cannot handle open-ended, multi-turn, multi-scenario exploratory queries like "Why has fish productivity declined in region X over the last 3 months?" [INFERENCE].
• Accuracy Potential: High for fixed queries, Zero for complex exploratory queries.
• Complexity: Medium.
• Explainability: High for rule matches, N/A for unhandled intents.
• Hackathon Feasibility: High.
• Production Feasibility: Low (rigid, brittle, fails user expectations for conversational intelligence). Confidence Tag: [INFERENCE]

Approach 3: Modular Multi-Agent Orchestration Engine with Spatial Analytics & Deterministic GIS Guardrails (Recommended Hybrid)

• Mechanism: A multi-agent network (using LangGraph/StateGraph) with specialized roles:
  1. Router/Supervisor Agent: Language detection, intent classification, task planning.
  2. Marine Data Discovery Agent: Queries STAC/ERDDAP/WMS catalogs for relevant rasters/vectors.
  3. Ocean & GIS Analytics Agent: Programmatically invokes PostGIS/xarray tools for spatial-temporal calculations (SST fronts, Chlorophyll overlap).
  4. Safety & Geofence Agent: Runs deterministic point-in-polygon checks against IMBL, storm alert buffers.
  5. Route Optimization Agent: Executes pathfinding on dynamic ocean wave cost grids.
  6. Synthesizer & Audio Agent: Formulates multi-lingual, explainable responses with map visualizations and text-to-speech.
• Problem Fit: High. Exactly mirrors the problem statement's call for collaborative specialized agents [PS].
• Accuracy Potential: High (AI handles planning/synthesis while deterministic code executes GIS and safety math).
• Complexity: High architectural design, modular implementation.
• Explainability: High (each agent records its intermediate state and evidence trace).
• Hackathon Feasibility: High (using modular Python services and mock state nodes).
• Production Feasibility: High (scalable, maintainable, secure). Confidence Tag: [INFERENCE]

10. Recommended Approach & Why

Selected Approach: Approach 3 (Modular Multi-Agent Orchestration Engine with Deterministic GIS Guardrails) [INFERENCE].

Justification

• Architectural Alignment: Directly fulfills ISRO's requirement for a multi-agent framework featuring autonomous task decomposition, specialized agent collaboration, and tool execution [PS].
• Safety-Critical Reliability: By separating llm-reasoning from gis-execution, deterministic calculations (e.g., "Is vessel within 2 nm of IMBL?") are handled by PostGIS/Shapely, completely eliminating LLM math hallucination [INFERENCE].
• Explainability: Each agent contributes evidence to a shared state graph. The Synthesizer Agent converts this audit log into explicit citations for the user (e.g., "PFZ calculated using SST front data from MOSDAC and Wave forecast from INCOIS OSF"), satisfying requirement [PS].

Why Alternatives Were Rejected

• Monolithic LLM (Approach 1): Unreliable for multi-step spatial tasks; calling 5 different GIS tools in a single context window leads to high latency, rate-limiting, and error cascades [INFERENCE].
• Rule-Based Bot (Approach 2): Fails on exploratory reasoning queries (Query g: "Why has productivity declined?") and multi-turn scenario building [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+-----------------------------------------------------------------------------------+
|                            PRESENTATION & INTERACTION LAYER                       |
|   +--------------------------+   +------------------------+   +-----------------+ |
|   | Progressive Web App (PWA)|   | High-Contrast UI Mode  |   | Multi-lingual   | |
|   | Leaflet/MapLibre WebGIS  |   | Voice Audio / Indic ASR|   | Offline Cache   | |
|   +--------------------------+   +------------------------+   +-----------------+ |
+-----------------------------------------------------------------------------------+
                                         |  HTTPS / WebSockets
+-----------------------------------------------------------------------------------+
|                        AGENT ORCHESTRATION LAYER (LangGraph)                      |
|                                                                                   |
|  +-----------------------------------------------------------------------------+  |
|  |                   SUPERVISOR / INTENT ROUTER AGENT                          |  |
|  | - Language Identification (IndicTrans2) | Multi-Turn Memory Management      |  |
|  | - Intent Classification & DAG Task Decomposition Plan                       |  |
|  +-----------------------------------------------------------------------------+  |
|                                        |                                          |
|       +--------------------------------+--------------------------------+         |
|       |                                |                                |         |
|  +----+------------------+    +--------+-----------------+    +---------+-------+  |
|  | DATA DISCOVERY AGENT  |    | OCEAN ANALYTICS AGENT    |    | SAFETY & GEOFENCE|  |
|  | STAC / ERDDAP Query   |    | SST/Chlorophyll Fronts   |    | PostGIS IMBL Check|  |
|  | Dataset Metadata Fetch|    | Temporal Trend Analysis  |    | Storm/Lightning   |  |
|  +-----------------------+    +--------------------------+    +-------------------+  |
|       |                                |                                |         |
|       +--------------------------------+--------------------------------+         |
|                                        |                                          |
|  +-------------------------------------+---------------------------------------+  |
|  | ROUTE OPTIMIZATION AGENT            | SYNTHESIZER & EXPLAINABILITY AGENT    |  |
|  | A* Pathfinding on SWH Cost Surface  | Multi-lingual Generation + Audio TTS  |  |
|  +-------------------------------------+---------------------------------------+  |
+-----------------------------------------------------------------------------------+
                                         |  Internal Python API Call
+-----------------------------------------------------------------------------------+
|                          DETERMINISTIC GIS & ENGINE TOOLKIT                       |
|   +-----------------------+   +----------------------+   +---------------------+  |
|   | GDAL/Rasterio Processing| | PostGIS Vector Engine|   | NetworkX Path Planner|  |
|   +-----------------------+   +----------------------+   +---------------------+  |
+-----------------------------------------------------------------------------------+
                                         |  Data Fetching / Ingestion
+-----------------------------------------------------------------------------------+
|                            DATA INGESTION & STORAGE LAYER                         |
|   +--------------------+   +-----------------------+   +-----------------------+  |
|   | PostgreSQL/PostGIS |   | Raster Tile Cache     |   | Vector Tile Server    |  |
|   | (Spatial DB)       |   | (COG / GeoTIFF Storage)|   | (Martin / Tippecanoe) |  |
|   +--------------------+   +-----------------------+   +-----------------------+  |
|                                        |                                          |
|   +------------------------------------+---------------------------------------+  |
|   | External Ingestors: MOSDAC (STAC), INCOIS (ERDDAP), IMD (CAP/REST)         |  |
|   +----------------------------------------------------------------------------+  |
+-----------------------------------------------------------------------------------+

Detailed Data-Flow Narrative

1. User Request Ingestion:

   • A fisherman near Rameswaram submits a Tamil voice query: "நாளை காலை கடலுக்குப் போவது பாதுகாப்பானதா? அருகில் மீன் பிடிக்கும் மண்டலம் எங்குள்ளது?" ("Is it safe to go to sea tomorrow morning? Where is the nearest fishing zone?") [INFERENCE].
   • PWA captures audio -> sends to Fast API backend -> processed by IndicSpeech ASR -> converted to Tamil text and translated to English internal representation [INFERENCE].

2. Supervisor Agent Task Decomposition:

   • Supervisor Agent identifies two sub-intents: Intent A (Sea Safety Check for next 24 hours), Intent B (Nearest PFZ discovery) [INFERENCE].
   • Creates execution DAG:
     • Step 1: Query Safety Agent for IMD warnings, wave heights (INCOIS OSF), and lightning.
     • Step 2: Query Ocean Analytics Agent for SST/Chlorophyll PFZ coordinates near current GPS.
     • Step 3: Run Safety Agent check on PFZ coordinates to ensure they are NOT inside IMBL or hazardous wave zones.
     • Step 4: Synthesize consolidated response.

3. Tool Execution & Safety Engine Verification:

   • Safety Agent: Queries PostGIS for active IMD warnings and wave rasters. Result: Wave height 1.8m (Safe < 2.5m), No lightning alert [INFERENCE].
   • Ocean Analytics Agent: Fetches latest EOS-6 OCM-3 Chlorophyll and GHRSST rasters -> runs Python overlap function -> identifies PFZ polygon 14 km East-Southeast [INFERENCE].
   • Geofence Tool: Checks PostGIS distance from PFZ to Sri Lanka IMBL boundary -> Distance is 6.2 Nautical Miles (Outside restricted buffer) [INFERENCE].

4. Response Synthesis & Multimodal Output Generation:

   • Synthesizer Agent formats multi-modal payload:
     • Text (Tamil): "நாளை காலை கடல் நிலை பாதுகாப்பானது (அலை உயரம் 1.8 மீ). உங்கள் படகில் இருந்து 14 கி.மீ தொலைவில் உள்ள சாத்தியமான மீன்பிடி மண்டலம் கண்டறியப்பட்டது. இலங்கை எல்லைக்கு 6.2 கடல் மைல் தொலைவில் பாதுகாப்பாக உள்ளது."
     • Audio: Converts Tamil text to spoken MP3 stream using IndicTTS [INFERENCE].
     • Map Layers: GeoJSON point for vessel, GeoJSON polygon for PFZ, red polyline for IMBL boundary, displayed on Leaflet map [INFERENCE].

12. Technology Stack

13. AI/ML Justification

Where AI Should NOT Be Used

• Geofencing & Boundary Violations: AI must NEVER predict whether a boat is across the International Maritime Boundary Line (IMBL). This MUST be executed deterministically using PostGIS spatial intersection (ST_Contains / ST_Distance) [INFERENCE].
• Marine Safety Thresholds: Storm warnings, wave height safety checks (e.g., SWH > 3.0m = Danger), and lightning proximity checks MUST rely on strict, deterministic threshold logic rather than LLM inference [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

+------------------------------------------------------------------------------------+
|                         LOCAL CLIENT STORAGE (PWA / MOBILE)                        |
|                                                                                    |
|  +------------------------------+     +-----------------------------------------+  |
|  | SQLite WASM / IndexedDB      |     | MBTiles / Vector Tile Store             |  |
|  | - Daily PFZ Vectors (GeoJSON) |     | - Compressed Coastal Base Map Tiles     |  |
|  | - Pre-fetched Weather Forecast |     | - IMBL & MPA Polyline Layers            |  |
|  | - User Vessel Profile & Logs |     | - Offline Visual Map Display            |  |
|  +------------------------------+     +-----------------------------------------+  |
+------------------------------------------------------------------------------------+
                                          ^  Synchronize on Connect
                                          |  (HTTP/2 JSON Compression)
+------------------------------------------------------------------------------------+
|                         CENTRAL SERVER ARCHITECTURE                                |
|                                                                                    |
|  +------------------------------+     +-----------------------------------------+  |
|  | PostgreSQL 16 / PostGIS      |     | Raster COG Storage (S3 / Local MinIO)   |  |
|  | - Master Geospatial Data     |     | - Cloud-Optimized GeoTIFFs (SST, Chl)   |  |
|  | - User Session History       |     | - Windowed spatial HTTP Range requests |  |
|  +------------------------------+     +-----------------------------------------+  |
+------------------------------------------------------------------------------------+

Low-Network & Offline Strategies

1. Cloud-Optimized GeoTIFF (COG) & Vector Tiles:
   • Remote raster files stored as COGs on server, allowing HTTP Range requests to read only the specific bounding box required for user query without downloading full multi-gigabyte satellite scenes [INFERENCE].
2. PWA Pre-Fetching Protocol:
   • Before departure from port (under Wi-Fi/4G), PWA automatically syncs next 48-hour OSF wave grids, PFZ vectors, and local IMBL layers for the vessel's home region (e.g., 50 km radius) into client-side IndexedDB [INFERENCE].
3. Bandwidth Reduction Techniques:
   • GeoJSON polylines compressed using Douglas-Peucker simplification algorithm before transmission [INFERENCE].
   • API responses serialized with Protocol Buffers / gzip JSON [INFERENCE].

15. Security, Privacy & Governance

1. Vessel Location & Operational Privacy:
   • Live GPS tracking data of fishing vessels is commercially sensitive (reveals secret fishing spots). GPS coordinates transmitted over HTTPS, encrypted at rest via AES-256 in PostgreSQL, and anonymized for analytics [INFERENCE].
2. IMBL & Boundary Geofencing Sensitivity:
   • Geofencing boundary shapefiles must match official Survey of India / Ministry of External Affairs definitions [INFERENCE]. System displays explicit UI disclaimers: "Geofence alerts are decision-support aids and do not replace official navigational equipment" [INFERENCE].
3. Role-Based Access Control (RBAC):
   • Public/Fishermen Role: Access to PFZ, weather, safety alerts, and route planning.
   • Researcher Role: Access to multi-temporal historical raster analytics and trend tools.
   • Authority Role: Access to broadcast alerts and spatial usage metrics [INFERENCE].
4. AI Output Safety & Governance:
   • Output guardrails check synthesized advisories against raw oceanographic thresholds before display to eliminate hallucinated safety claims [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (Must Build for Hackathon Demo)

• Core Multi-Agent Pipeline: Supervisor Router Agent + Data Discovery Agent + Ocean Analytics Agent + Safety & Geofencing Agent + Synthesizer Agent running on LangGraph [INFERENCE].
• Natural Language Conversational Interface: React/Next.js Web App supporting text and voice input (English, Hindi, Tamil) [INFERENCE].
• Spatial Reasoning over SST & Chlorophyll: Programmatic overlay of mock/cached SST thermal fronts and Chlorophyll-a layers to identify PFZ coordinates [INFERENCE].
• Interactive Geospatial Visualizer: MapLibre GL map displaying PFZ polygons, wave height heatmaps, user location, and IMBL warning lines [INFERENCE].
• Proactive Geofencing Alert Engine: Real-time point-in-polygon engine triggering visual/audio warnings when mock vessel approaches within 5 km of IMBL [INFERENCE].
• Safe Route Pathfinding: A-star algorithm generating path avoiding high wave height zones [INFERENCE].

Described Only / Out of Scope for Hackathon

• Native hardware integration with NAVIC satellite messaging transponders [INFERENCE].
• Full production ingestion of all 10 coastal state languages [INFERENCE].
• Multi-year petabyte-scale historical satellite archival processing backend [INFERENCE].

17. Live Demo Simulation Plan

To ensure a seamless, reliable presentation during the SIH finals without dependence on flaky live government server endpoints or active sea conditions, the team will implement a transparent Simulation Engine:

+-----------------------------------------------------------------------------------+
|                            DEMO SIMULATION PIPELINE                               |
|                                                                                   |
|  +-------------------------+     +------------------------+   +-----------------+ |
|  | Cached Real Datasets    |     | Vessel Movement        |   | Weather Hazard  | |
|  | - INCOIS PFZ Sample GeoJSON   | Simulator              |   | Injector        | |
|  | - EOS-6 OCM-3 COG File  |     | - Interactive GPS Track|   | - Toggle High   | |
|  | - GEBCO Bathymetry Tile |     |   Playback in Map      |   |   Waves / Storm | |
|  +-------------------------+     +------------------------+   +-----------------+ |
|                 |                            |                        |           |
|                 +----------------------------+------------------------+           |
|                                              |                                    |
|                                              v                                    |
|                          +---------------------------------------+                |
|                          | UNIFIED ORCA SIMULATION CONTROLLER    |                |
|                          | - Preserves full real data flow       |                |
|                          | - Clearly flags "SIMULATED FEED" in UI|                |
|                          +---------------------------------------+                |
+-----------------------------------------------------------------------------------+

1. Simulated Vessel GPS Feed:
   • Interactive UI slider allowing judges to drag a mock vessel across the Palk Strait towards the IMBL boundary to demonstrate real-time geofence audio alerts [INFERENCE].
2. Weather Hazard Injector:
   • Demo panel button to inject a high wave alert (SWH = 3.8m) or IMD lightning strike alert near the vessel to demonstrate dynamic route recalculation [INFERENCE].
3. Data Authenticity:
   • All sample rasters (SST, Chlorophyll) derived from actual archived EOS-6 and Copernicus passes over the Bay of Bengal / Arabian Sea [RESEARCHED]. Simulated feeds explicitly labelled in UI as "Simulated Real-Time Stream (Data Source: Verified MOSDAC/INCOIS Archive)" [INFERENCE].

18. Implementation Roadmap

Phase 1: Architecture Setup & Ingestion Pipeline
[01 Nov 2025 - 31 Jan 2026]
├── Set up PostgreSQL/PostGIS & Qdrant databases
├── Ingest static boundaries (IMBL, MPAs) & GEBCO Bathymetry
└── Implement COG raster loader & ERDDAP wrapper APIs

Phase 2: Agent Framework & Tool Development
[01 Feb 2026 - 30 Apr 2026]
├── Build LangGraph state graph (Supervisor, Data, Safety Agents)
├── Implement Python deterministic tools (PostGIS distance, xarray overlap)
└── Implement A* pathfinding algorithm on wave cost surfaces

Phase 3: Multi-Lingual NLP & Frontend WebGIS
[01 May 2026 - 31 Jul 2026]
├── Integrate IndicTrans2 & AI4Bharat Speech ASR/TTS
├── Build React/MapLibre GL dashboard & PWA offline cache
└── Connect agent state outputs to map layer rendering

Phase 4: Integration, Field Testing & SIH Polish
[01 Aug 2026 - 30 Sep 2026]
├── Run false-positive & offline stress testing
├── Implement Live Demo Simulator Controller
└── Complete SIH documentation & video submission (Deadline: 30 Sep 2026)

19. Testing & Validation Plan

1. Functional Agent & Tool Testing:
   • Test intent router across 200+ multi-lingual sample prompts covering queries a-h [INFERENCE]. Verify DAG task execution success rate (>95%) [INFERENCE].
2. Geofencing & Safety False-Positive / False-Negative Testing:
   • Test vessel points at varying distances from IMBL (10km, 5km, 1km, 0.1km across boundary) [INFERENCE]. Guarantee 0% False Negatives (never fail to alert when vessel crosses boundary) [INFERENCE].
3. AI Hallucination Verification:
   • Pass out-of-bounds or nonsensical queries (e.g., "Where is PFZ in the Thar Desert?") [INFERENCE]. System must gracefully decline and explain geographic constraints [INFERENCE].
4. Offline PWA Performance Testing:
   • Simulate 0kbps network in Chrome DevTools [INFERENCE]. Verify that pre-cached map tiles, local IndexedDB PFZ lines, and audio alerts function correctly offline [INFERENCE].

20. Risk Register

21. Cost & Scalability

Hackathon Architecture (Minimal Cost)

• Hosting: Free-tier / low-cost cloud VM (e.g., Hetzner / AWS T4 GPU instance ~$50/month) [INFERENCE].
• LLM Compute: Local quantized Llama-3-8B model or pay-as-you-go GPT-4o API (<$20 total testing) [INFERENCE].
• Database: Self-hosted PostgreSQL/PostGIS + Qdrant in Docker containers ($0) [INFERENCE].
• Total Hackathon Cost: ~$50 - $100 [INFERENCE].

Production System Deployment (ISRO / INCOIS Scale)

• Infrastructure: Kubernetes Cluster (EKS/GKE or On-Premise ISRO High-Performance Compute HPC) with autoscaling GPU nodes for vLLM inference [INFERENCE].
• Storage: MinIO / AWS S3 for multi-terabyte Cloud-Optimized GeoTIFF (COG) satellite archives [INFERENCE].
• CDNs & Map Services: CDN tile edge caching (Cloudflare) for low-latency vector/raster tile distribution to millions of coastal users [INFERENCE].

22. Differentiation

23. Future Scope

1. Direct NAVIC Satellite Broadcast Integration: Partnering with ISRO to encode compressed safety alerts into NAVIC messaging receivers installed on small fishing vessels for beyond-cellular messaging [INFERENCE].
2. On-Vessel Edge AI Processing: Deploying lightweight quantized edge-agents on low-cost Raspberry Pi / Jetson devices onboard vessels for local sonar/fish-finder multi-modal fusion [INFERENCE].
3. Automated Marine Heatwave & Ecological Anomaly Detection: Expanding agent tools to continuously monitor long-term trends in coral bleaching, harmful algal blooms (HAB), and coastal erosion [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions [ASSUMPTION]

• [ASSUMPTION] A 4-6 student team can utilize pre-trained multi-lingual open models (e.g., IndicTrans2, Whisper) and open-source agent frameworks (LangGraph) within the project timeline.
• [ASSUMPTION] Historical open oceanographic data from Copernicus Marine and NOAA GHRSST can serve as reliable proxies during development when live ISRO MOSDAC endpoints undergo maintenance.
• [ASSUMPTION] Target users possess basic Android smartphones capable of running modern WebPWAs and HTML5 audio/geolocation APIs.

Consolidated Open Items & Validation Requirements [NEEDS VALIDATION]

• [NEEDS VALIDATION] Exact real-time API authentication protocols and rate limits for official live ISRO MOSDAC STAC catalog and INCOIS ERDDAP services.
• [NEEDS VALIDATION] Formal legal specifications for Survey of India maritime boundary polylines for exact high-precision IMBL compliance.
• [NEEDS VALIDATION] Field testing feedback from local coastal fishing communities regarding dialect accuracy of AI4Bharat text-to-speech engines in real-world marine noise environments.

Final Recommendation

Adopt Approach 3 (Modular Multi-Agent Orchestration Engine with Deterministic GIS Guardrails). Build a state-driven LangGraph orchestrator backed by PostGIS spatial analytics, Cloud-Optimized GeoTIFF raster processing, and an offline-first PWA frontend with multi-lingual voice support [INFERENCE].

Recommended Stack

• Frontend: React / Next.js PWA + MapLibre GL JS
• Backend API: Python FastAPI
• Agent Framework: LangGraph (StateGraph)
• Database: PostgreSQL 16 + PostGIS 3.4
• Spatial Processing: Rasterio, xarray, Shapely, GDAL
• AI / Speech: Quantized Llama-3-70B (vLLM) + AI4Bharat IndicTrans2 & IndicTTS / Whisper

Biggest Dependency

Access to reliable, structured oceanographic raster streams (SST, Chlorophyll) and ocean state forecasts (waves, currents) [INFERENCE].

Biggest Risk

Cellular connectivity drop-outs at sea preventing live API calls (Mitigated via PWA pre-fetching and IndexedDB local storage) [INFERENCE].

Biggest Technical Challenge

Programmatic, hallucination-free invocation of spatial raster correlation tools by AI agents for dynamic queries [INFERENCE].

First 5 Actions

1. Initialize Git repository and set up PostgreSQL with PostGIS extension [INFERENCE].
2. Download and seed static vector layers (IMBL boundary, Marine Protected Areas, coastal landing centers) into PostGIS [INFERENCE].
3. Implement Python FastAPI microservice wrapping Rasterio/xarray for SST and Chlorophyll multi-criteria raster overlay [INFERENCE].
4. Build the core LangGraph state graph with Supervisor, Safety, and Ocean Analytics agents [INFERENCE].
5. Develop the MapLibre GL frontend dashboard with multi-lingual text/speech input support [INFERENCE].

Validation Before Coding

• Verify local PostGIS spatial query execution times for point-in-polygon checks (<10ms target) [INFERENCE].
• Validate sample NetCDF/GeoTIFF ingestion from open ERDDAP/Copernicus endpoints [INFERENCE].
• Test LangGraph tool-calling accuracy with quantized local LLM models [INFERENCE].

Buildability Score /100

88/100 (Highly feasible for a skilled 4-6 student team using modern Python agent frameworks, PostGIS, and open-source GIS libraries within the timeline ending 30 September 2026) [INFERENCE].

Confidence

HIGH (The solution design strictly adheres to explicit PS requirements while using proven open-source geospatial and agentic architectures) [INFERENCE].

Research Sources

1. Potential Fishing Zone (PFZ) Advisory Services - ESSO-INCOIS

   • URL: https://incois.gov.in/MarineFisheries/PfzAdvisory [RESEARCHED]
   • Source Type: Official Government Portal (Ministry of Earth Sciences, Govt. of India).
   • Supported Claim: Explains PFZ generation methodology using satellite SST and chlorophyll data across 586 landing centers.
   • Reliability: High / Authoritative.

2. SAMUDRA Mobile Application - ESSO-INCOIS

   • URL: https://incois.gov.in/site/SAMUDRA/index.html [RESEARCHED]
   • Source Type: Official Government Application Documentation.
   • Supported Claim: Identifies existing features (PFZ, ocean state forecasts, high wave alerts, tsunami warnings, regional language support).
   • Reliability: High / Authoritative.

3. MOSDAC & VEDAS Earth Observation Services - ISRO

   • URL: https://www.isro.gov.in/EOS_06.html [RESEARCHED]
   • Source Type: Official Government Space Agency Documentation (ISRO).
   • Supported Claim: Details EOS-6 (Oceansat-3) payloads including OCM-3 (Ocean Color Monitor) and SSTM-1 (Sea Surface Temperature Monitor).
   • Reliability: High / Authoritative.

4. Bhoonidhi API & Data Access Specifications - NRSC / ISRO

   • URL: https://bhoonidhi.nrsc.gov.in/bhoonidhi-api/ [RESEARCHED]
   • Source Type: Official Government Data Hub API Specification.
   • Supported Claim: Confirms STAC catalog specification and programmatic API access protocols for ISRO satellite data products.
   • Reliability: High / Authoritative.

5. India Meteorological Department (IMD) API Portal

   • URL: https://api.imd.gov.in/public/api_reference.html [RESEARCHED]
   • Source Type: Official Government Weather API Service Documentation.
   • Supported Claim: Lists available endpoints for coastal bulletins, cyclone warnings, radar, and lightning strike data.
   • Reliability: High / Authoritative.