From fragmented marine data to intelligent, constraint-aware decisions.

ORCA is a modular multi-agent marine intelligence platform developed for Smart India Hackathon 2026 — Problem Statement SIH26176. It combines conversational AI, specialized marine AI/ML models, geospatial reasoning, multi-source data integration, and deterministic safety constraints to turn complex marine information into actionable, evidence-backed recommendations.

Team: Code Newbies_TS16
Team ID: 152277
Problem Statement: SIH26176 — ORCA Marine EcOsystem Reasoning with Collaborative Agents
Theme: Space Technology
Category: Software

Marine decision-making is difficult because useful information is fragmented across different sources, formats, models, and operational systems.

ORCA is designed around five key challenges:

• Scattered marine data — information is distributed across multiple marine, weather, satellite, and geospatial sources.
• High operational risk — changing weather, strong currents, waves, and natural hazards can rapidly affect maritime operations.
• Limited actionable insight — raw environmental data is difficult to translate into practical decisions.
• Poor connectivity at sea — intermittent connectivity makes lightweight, accessible interfaces important.
• Language and literacy barriers — marine users may communicate through different local languages and conversational forms.

ORCA follows a reasoning pipeline that converts a natural-language request into a coordinated, constraint-aware decision:

User Query
    │
    ▼
┌───────────────────────┐
│  1. ASK               │
│  Natural-language     │
│  voice / text input   │
└───────────┬───────────┘
            ▼
┌───────────────────────┐
│  2. DATA INTEGRATION  │
│  Multi-source marine  │
│  and environmental    │
│  information          │
└───────────┬───────────┘
            ▼
┌───────────────────────┐
│  3. UNDERSTAND        │
│  Intent, location,    │
│  time, activity and   │
│  context extraction   │
└───────────┬───────────┘
            ▼
┌───────────────────────┐
│  4. ANALYZE           │
│  Weather, ocean, PFZ, │
│  risk & geospatial    │
│  reasoning            │
└───────────┬───────────┘
            ▼
┌───────────────────────┐
│  5. VALIDATE & DECIDE │
│  Safety rules and     │
│  operational          │
│  constraints          │
└───────────┬───────────┘
            ▼
┌───────────────────────┐
│  6. RESPOND           │
│  Localized, actionable│
│  evidence-backed      │
│  recommendation       │
└───────────────────────┘

The system is intended to combine multiple marine information sources and specialized agents rather than relying on a single general-purpose model for every decision.

The ORCA platform is organized into separate layers so that the API/backend, reasoning engine, domain intelligence, models, and deployment infrastructure can evolve independently.

flowchart TB
    U[Users & Clients]

    subgraph CLIENTS["User & Client Layer"]
        WEB[Web Application]
        MOBILE[Mobile Application]
        CC[Command Center / Dashboard]
        RA[Researchers & Authorities]
    end

    subgraph API["API & Security Layer"]
        FASTAPI[FastAPI Backend]
        AUTH[Authentication & Authorization]
        RATE[Rate Limiting]
        VALID[Input Validation]
        ROUTE[Request Routing]
    end

    subgraph ORCA["ORCA Multi-Agent Intelligence Layer"]
        ROUTER[Conversational Router<br/>Qwen]
        ORCH[Multi-Agent Orchestrator]
        PLAN[Query Understanding & Task Planning]
        SELECT[Dynamic Agent Selection]
        CONTEXT[Context Management]
        SAFETY[Safety & Guardrails]
        SYNTH[Response Synthesis]
    end

    subgraph AGENTS["Domain Agents"]
        WEATHER[Weather Agent]
        OCEAN[Ocean / Tide Agent]
        PFZ[PFZ Agent]
        PRODUCTIVITY[Productivity Agent]
        RISK[Risk & Safety Agent]
        GEO[Geospatial Agent]
        REC[Recommendation Agent]
    end

    subgraph DATA["External / Official Data Sources"]
        INCOIS[INCOIS]
        IMD[IMD]
        MOSDAC[MOSDAC]
        OPENMETEO[Open-Meteo]
        GEO_DATA[Natural / Geospatial Data]
    end

    subgraph ML["ML Models Layer"]
        QWEN[Qwen / Conversational Model]
        XGB[XGBoost Models]
        LSTM[LSTM / Keras Models]
    end

    subgraph INFRA["Deployment & Infrastructure"]
        DOCKER[Docker]
        K8S[Kubernetes]
        HOST[Self-Hosted / Server]
        NGROK[ngrok]
        CI[GitHub Actions]
    end

    U --> WEB
    U --> MOBILE
    U --> CC
    U --> RA

    WEB --> FASTAPI
    MOBILE --> FASTAPI
    CC --> FASTAPI
    RA --> FASTAPI

    FASTAPI --> AUTH
    AUTH --> RATE
    RATE --> VALID
    VALID --> ROUTE

    ROUTE --> ROUTER
    ROUTER --> ORCH

    ORCH --> PLAN
    ORCH --> SELECT
    ORCH --> CONTEXT
    ORCH --> SAFETY
    ORCH --> SYNTH

    ORCH --> AGENTS

    WEATHER --> IMD
    WEATHER --> OPENMETEO
    OCEAN --> OPENMETEO
    PFZ --> INCOIS
    GEO --> GEO_DATA
    AGENTS --> ML

    FASTAPI --> HOST
    HOST --> DOCKER
    DOCKER --> K8S
    HOST --> NGROK
    HOST --> CI

Implementation note: The two major repositories currently documented in the ORCA organization have distinct responsibilities. backend provides the HTTP/API layer, authentication, sessions, rate limiting, and Supabase integration, while agent-orchestration contains the conversational routing, DAG-based orchestration, domain agents, marine ML inference, and safety logic.

The API gateway component of ORCA.

Its responsibilities include:

• FastAPI HTTP endpoints
• JWT authentication
• Rate limiting
• In-memory conversational session management
• Supabase/PostgreSQL access
• Request validation
• Backend observability and latency reporting
• Delegation of natural-language queries to the orchestration engine

The backend currently exposes verified endpoints including:

GET  /api/v1/health
POST /api/v1/orca/query
POST /api/v1/orca/command-center/query
GET  /api/v1/orca/sessions/{session_id}/history

The command-center endpoint is a prototype endpoint and is disabled by default.

The reasoning and multi-agent intelligence component.

Its responsibilities include:

• Qwen-based conversational routing
• Intent, location, temporal and activity extraction
• QueryPlan construction
• DAG-based agent orchestration
• Concurrent execution of domain agents
• Marine risk assessment
• PFZ analysis
• Weather and ocean analysis
• Geospatial reasoning
• Safety rules and constraint arbitration
• Recommendation synthesis
• Multilingual response generation

The repository currently provides a standalone CLI interface through chat.py.

ORCA decomposes marine reasoning into specialized domain agents rather than requiring one model to perform every task.

The orchestration engine resolves dependencies between agents and executes the required subset according to the query plan.

ORCA combines conversational AI with specialized predictive models.

The current agent-orchestration implementation contains the domain-specific model inference, while backend delegates reasoning requests to that layer.

ORCA is designed around multi-source marine intelligence.

The current repository implementations distinguish between data sources that are directly consumed by a component and sources accessed by delegated domain agents.

For example, backend directly uses Supabase for database-backed state and observations, while the live INCOIS/IMD/Open-Meteo integrations are handled by the domain/orchestration layer.

A central design principle of ORCA is that conversational AI should not be the sole authority for safety-critical recommendations.

The reasoning flow separates interpretation from validation:

Natural-Language Query
        ↓
LLM-based Understanding
        ↓
Structured Query Plan
        ↓
Relevant Domain Agents
        ↓
Environmental & Marine Evidence
        ↓
Risk Assessment
        ↓
Deterministic Safety Rules
        ↓
Constraint Arbitration
        ↓
Final Recommendation

This enables ORCA to combine learned models with explicit operational constraints.

Where catastrophic conditions are detected, the safety layer can override an otherwise favorable recommendation and issue a safety-oriented response such as RETURN TO HARBOR.

ORCA is designed to support questions such as:

"Is it safe to go fishing near Digha tomorrow?"

"Where is the nearest Potential Fishing Zone today?"

"Which fishing areas should be avoided because of hazardous marine conditions?"

"What is the safest route for a fishing vessel considering
weather and sea-state conditions?"

"Why has fish productivity declined in this coastal region?"

"I want to go on a fishing trip near Kochi. What areas
should I consider?"

The actual agents selected and executed depend on the query's intent, location, time and required reasoning.

ORCA is designed to provide more than a final label or recommendation.

The response can include supporting information such as:

• Weather conditions
• Ocean state
• Wave / wind conditions
• Marine risk
• PFZ information
• Geospatial constraints
• Official warning/advisory information
• Agent-level evidence
• Reasoning behind the resulting recommendation

This is intended to make recommendations inspectable rather than presenting an unexplained output.

The ORCA technical approach supports conversational interaction in:

• English
• Bengali
• Hinglish

The orchestration layer extracts intent and context from natural-language queries and generates localized responses.

The SIH solution concept also emphasizes voice/text interaction and accessibility for users operating with limited connectivity or different levels of technical literacy.

The SIH technical architecture proposes a modular deployment model that can support:

• Docker-based containerization
• Kubernetes orchestration
• Self-hosted or server-based deployment
• GitHub Actions for CI/CD
• ngrok for exposing development services
• Web and mobile clients

The current repositories should be distinguished from the broader deployment architecture shown in the SIH presentation: some deployment elements represent the proposed/scalable architecture rather than functionality contained in the two repositories documented here.

The SIH presentation compares ORCA's intended capabilities with existing marine/weather systems and highlights areas such as:

• Uncertainty quantification
• Real-time data assimilation
• Domain-specific data processing
• Spatial-temporal reasoning
• Specialized AI/ML
• Multi-source fusion
• Natural-language interaction
• Explainable recommendation
• Multi-agent orchestration
• Marine risk assessment
• Geofencing
• Route optimization

The presentation also cites research covering machine learning for ocean data assimilation, probabilistic weather forecasting, data assimilation for ocean forecasting, OceanGPT, medium-range global weather forecasting, FourCastNet, and vessel safety/risk systems.

References included in the SIH presentation:

1. D. Grande, R. Buizza, and A. Storto, Machine learning in ocean data assimilation: Advances, gaps and the road to operations, Ocean Modelling, vol. 200, 102676, 2026.
2. I. Price, A. Sanchez-Gonzalez, F. Alet, et al., Probabilistic weather forecasting with machine learning, Nature, vol. 637, pp. 84–90, 2025.
3. M. J. Martin, L. Floet, I. Bertino, and A. M. Moore, Data assimilation schemes for ocean forecasting: State of the art, State of the Planet, vol. 5, 2025.
4. Z. Bi, N. Zhang, Y. Xue, Y. Du, D. Ji, G. Zheng, and H. Chen, OceanGPT: A Large Language Model for Ocean Science Tasks, Proceedings of ACL, 2024.
5. K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian, Accurate medium-range global weather forecasting with 3D neural networks, Nature, vol. 619, pp. 533–538, 2023.
6. J. Pathak, S. Subramanian, P. Harrington, et al., FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators, 2022.
7. T. M. Balakrishnan Nair, et al., Development of small vessel advisory and forecast services system for safe navigation and operations at sea, Journal of Operational Oceanography, 2020.

The proposed ORCA architecture is designed around:

• Modular architecture
• Independent agent integration
• Python-based model integration
• Existing open/government data sources

• Trusted open and government datasets
• Real-time data availability where supported
• Multi-source marine information

• Open-source technologies
• Existing infrastructure
• Cost-conscious deployment

• API-first integration
• Web/mobile accessibility
• Modular backend architecture

The SIH proposal identifies potential value across:

• Fishermen — fishing-zone intelligence and safety information
• Authorities — timely alerts and data-driven decision support
• Businesses — operational and route optimization
• Coastal communities — early information and preparedness
• Researchers & students — unified access to multi-source marine information

Smarter Decisions. Safer Seas. Stronger Communities.

The proposed impact areas include:

• Safer and more informed fishing decisions
• Early awareness of marine hazards
• More efficient maritime operations
• Better route and spatial planning
• Improved access to integrated marine information
• Support for coastal preparedness
• Easier access to complex environmental information
• A foundation for further marine research and decision-support systems

These represent the intended impact of the ORCA solution; actual operational outcomes depend on deployment, data availability, validation, and user adoption.

The organization is structured around separate components rather than placing the complete system into a single repository.

Additional repositories/components can be added to this map as the ORCA platform evolves.

The organization contains a working prototype architecture, but the current repositories should be distinguished from the complete long-term architecture presented in the SIH proposal.

In particular:

• backend currently provides the HTTP/API layer and depends on the sibling agent-orchestration repository.
• agent-orchestration currently provides the standalone CLI reasoning/orchestration interface.
• Some external providers are configurable or delegated to the orchestration layer.
• The backend currently uses in-memory session state with Supabase persistence support.
• The orchestration layer uses local ML models and external marine data integrations.
• The complete production-scale deployment architecture shown in the SIH presentation is a broader target architecture rather than a claim that every component is already contained in these two repositories.

This distinction keeps the organization documentation aligned with the implemented repositories while preserving the broader ORCA system vision.

Code Newbies_TS16

Smart India Hackathon 2026
Problem Statement: SIH26176
Team ID: 152277

Marine EcOsystem Reasoning with Collaborative Agents

From fragmented marine data to intelligent, constraint-aware decisions.