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 GitHubSistema 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-tower1 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
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.

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.

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

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.

Agent Pipeline 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 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 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 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 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 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.

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.
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.