Skip to content

Architecture Overview

The Fleet Decision Platform follows a modular, pipeline-based architecture designed for scalability and maintainability.

Design Principles

Principle Description
Modular Each capability is a separate, replaceable module
Config-Driven All parameters externalized to YAML/JSON
Pipeline-Based Clear data flow between stages
API-First Everything accessible via REST API
Explainable Transparency at every decision point

High-Level Architecture

graph TB
    subgraph External["External Data"]
        NYC[NYC Taxi Data]
        NASA[NASA Turbofan]
    end

    subgraph DataLayer["Data Layer"]
        RAW[(Raw Data)]
        PROC[(Processed)]
    end

    subgraph MLLayer["ML Layer"]
        FORE[Forecasting]
        RISK[Risk Scoring]
    end

    subgraph OptLayer["Optimization Layer"]
        CONST[Constraints]
        OPT[Optimizer]
    end

    subgraph APILayer["API Layer"]
        API[FastAPI]
        DOCS[Documentation]
    end

    External --> DataLayer
    DataLayer --> MLLayer
    MLLayer --> OptLayer
    CONST --> OptLayer
    OptLayer --> APILayer

Layer Details

  • Data Layer

    Handles data ingestion, preprocessing, and storage.

    Data Flow

  • ML Layer

    Demand forecasting and risk prediction models.

    Module Design

  • Optimization Layer

    Cascading optimization with constraints.

    System Overview

  • API Layer

    REST API endpoints and documentation.

    API Reference

Technology Stack

Layer Technologies
Language Python 3.9+
ML XGBoost, scikit-learn, (Prophet, PyTorch)
Optimization OR-Tools, (PuLP)
API FastAPI, Pydantic, Uvicorn
Data Pandas, NumPy, PyArrow/Parquet
Database PostgreSQL, Redis, (Pinecone)
Documentation MkDocs Material

Phased Implementation

timeline
    title Development Roadmap

    Phase 1 : MVP
             : XGBoost forecasting
             : Min-cost flow optimization
             : Basic API

    Phase 2 : Enhanced ML
             : Hierarchical forecasting
             : Risk classification
             : SHAP explainability

    Phase 3 : Contract Intelligence
             : OCR/NLP extraction
             : Constraint registry
             : RAG search

    Phase 4 : Production
             : MILP refinement
             : Survival analysis
             : Full deployment

Next Steps