convez.ai
AgenteAutomaçãoAvançado
sarthakshirsatai-lab

Control Tower de Logística com IA Agêntica

sarthakshirsatai-lab/ai-logistics-control-tower

6 agentes coordenados para automação de exceções em logística de última milha

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema de 6 agentes IA que simula tomada de decisão operacional em logística de última milha europeia (Alemanha, França, Holanda). Detecta exceções de entrega, monitora performance de couriers, recomenda resoluções e comunica com clientes via IA generativa. Implementado com Claude Code, Python e Flask com loop human-in-the-loop.

Instalação

$npx convez-ai sarthakshirsatai-lab/ai-logistics-control-tower

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Detecção automática de exceções em entregas B2C
  • Scoring e monitoramento de performance de couriers
  • Recomendação inteligente de resoluções com priorização SLA/custo/CX
  • Comunicação automática com clientes via WhatsApp/email

Veredito

Projeto educacional robusto que demonstra orquestração real de múltiplos agentes em um caso de uso operacional concreto. A arquitetura human-in-the-loop é prática, mas documentação e generalização para outros domínios carecem de detalhamento.

README

A tradução para o português deste README ainda está na fila.

AI Logistics Control Tower

A personal learning project — building agentic AI systems that simulate operational decision-making in logistics, based on 9.5 years of experience in end-to-end logistics operations, product building and D2C brand building across Nigeria and India.

Built with Claude Code | Python | Flask | Anthropic Claude API | Pinecone Vector DB


🎥 Watch Demo

Watch the Working Prototype on LinkedIn


What This Is

A 6-agent agentic AI system with a functional Flask web application that simulates how AI can support operational decision-making in European last-mile delivery — across exception detection, courier performance monitoring, resolution recommendations, customer communication, orchestration, and natural language querying.

Each agent solves a distinct operational problem. Each builds on the output of the previous one. A Dispatch Controller (human) reviews escalated decisions via the web app — this is human-in-the-loop AI in practice.


Working Prototype — Screenshots

Dashboard

6 agents. One control tower. The dashboard shows real-time pipeline output — exceptions detected, auto-handled vs escalated, and agent-by-agent run summary.

Dashboard


Courier Performance

30-day rolling scorecard across 4 couriers — RAG rating (Red / Amber / Green) with recommended action. Zipovva Exxpress flagged for suspension. Customer absent and address errors excluded from courier scoring — not courier fault.

Courier Performance


Delivery Issues

Full exception log — Shipment ID, Courier, Country, Issue Type, Urgency, Order Value, Status and Action Taken. Filterable by status.

Delivery Issues


SQL Query Assistant

Plain English querying on top of the SQLite database. Any logistics professional can ask questions without knowing SQL — and get instant answers. Self-correction retry loop fires automatically if the first SQL query fails or returns empty.

SQL Query Assistant


Agent Pipeline Flow

Orchestrator Agent Flow


The 6 Agents

Agent 1 — Last Mile Exception Detector

Monitors B2C last mile shipments across Germany, France and Netherlands. Detects delivery exceptions and classifies them by type (failed_attempt, customer_absent, address_error) and urgency (HIGH / MEDIUM / LOW).

Agent 1 Simulation Results


Agent 2 — Courier Performance Monitor

Analyses 30 days of shipment data across 440 shipments. Scores each courier on On-Time Delivery %, courier-attributable miss rate and exception rate. Outputs RAG rating and recommended action per courier.

Agent 2 Scorecard


Agent 3 — Exception Resolution Agent

Given a shipment exception, scores all available resolution options using a composite formula (SLA × 0.40 + Cost × 0.35 + CX × 0.25). Determines whether to AUTO-EXECUTE or escalate to the Dispatch Controller.

Agent 3 Exception Resolution


Agent 4 — Customer Communication Agent

Generates empathy-driven customer messages via Anthropic Claude API. Two-layer output: structured template (Layer 1) + AI-generated empathy message (Layer 2). Channel: WhatsApp (≤€150 orders) | Email + WhatsApp (>€150 orders).

Agent 4 Customer Communication


Agent 5 — Orchestrator + Flask Web Application

Orchestrates all 4 agents sequentially. Serves a full Flask web application with 5 screens: Dashboard | Delivery Issues | Needs Your Decision | Courier Performance | Decision History. Human-in-the-loop: Dispatch Controller can APPROVE / REJECT / MODIFY escalated decisions.

Agent 5 Orchestrator


Agent 6 — SQL Query Assistant

Enables any logistics professional to query the SQLite database in plain English — no SQL knowledge needed. One Claude API call converts the question to SQL, executes it locally, and returns a plain English answer. Features a self-correction retry loop: if the first SQL query fails or returns empty, the agent retries automatically with the error context included.

Architecture:

  • Plain English question → Claude API → SQL query generated
  • SQL runs locally against SQLite database
  • Result formatted in plain English — no second API call
  • Data never leaves the machine after the SQL generation step
  • Self-correction loop: retry with error context on failure

Sample outputs:

  • "Which courier had the most failures?" → PostNL — 10 courier-attributable failures
  • "What is the OTD rate for PostNL?" → 95.0%
  • "How many shipments were handled by DHL?" → 34

Cost: ~$0.001 per query — roughly ₹0.10 per question

Production note: In this prototype, the database schema is shared with the AI to generate queries. In production, deploy on AWS Bedrock or Azure AI — so nothing leaves the enterprise environment. The logic is the same. The infrastructure is what changes.

Agent 6 SQL Query Assistant


Tech Stack

🔹 Frontend: HTML + CSS + JavaScript 🔹 Backend: Python + Flask 🔹 AI Layer: Anthropic Claude API (claude-sonnet-4-5) 🔹 Agent Framework: Custom 6-agent pipeline (no LangChain) 🔹 Data: SQLite 🔹 Built with: Claude Code — plain English prompts only


Geography & Couriers

Countries: Germany · France · Netherlands Couriers: DHL · PostNL · DPD · Zipovva Exxpress (fictional)


Built By

Sarthak Shirsat Founder, Acharooz | Ex-Movam | Ex Tolaram Group | IIM Mumbai MBA 9.5 years across fleet management, last mile logistics, B2B logistics SaaS and D2C brand building.

LinkedIn


Disclaimer

All shipment data is simulated. DHL, PostNL and DPD are referenced for simulation realism only — no affiliation. Zipovva Exxpress is a fictional courier. Built with Claude Code.

Ver o original no GitHub

Bom para

Relacionados