Background

As the market leader in automotive aftermarket services, the organization manages a highly complex cross border supply chain, handling long lead time imported parts and high frequency island wide distribution. While basic digital forecasting tools were introduced to enhance efficiency, the sheer scale and complexity of operations meant that relying solely on passive data monitoring was no longer sufficient for the dynamic logistics environment. Frontline staff faced a critical challenge: existing dashboards provided only historical stock snapshots but failed to simulate the ripple effects of complex decisions such as optimizing order quantities and cross docking under tight capacity constraints. This gap between static tools and real world demands led to decision latency, making it difficult to strike a precise balance between the risks of stockouts and excess inventory. To break this bottleneck, the organization needed to upgrade its operational core from viewing reports to an autonomous decision engine.

Problem

Decision Latency and Data Silos Despite having access to massive amounts of data, frontline operators suffered from severe Decision Latency. Current dashboards could show what is missing (passive monitoring) but could not simulate what decision solves the problem (active planning). Staff had to rely on fragmented information across disparate systems and use intuition to decide whether to expedite orders or initiate inter-warehouse transfers, without the ability to instantly assess the specific impact on logistics costs or service levels. This reliance on experience based decision making often resulted in a lose lose situation: inflated inventory days or compromised service availability.

Solution

Building a Deterministic Autonomous Decision Hub To overcome the limitations of static reporting and achieve autonomous decision making, I designed a 5 Layer Multi Agent System acting as an intelligent decision hub on top of the existing ERP. The core of this solution is the Manager Agent, which autonomously orchestrates complex logistics workflows by decomposing daily operational challenges into executable planning tasks.

On the data and protocol layers, acknowledging the zero tolerance policy for numerical errors in inventory management, I engineered a GraphRAG Engine with Rich Node Anchoring and developed a Text to Cypher Interface. This innovation serves as the bridge between agents and the enterprise database, ensuring agents can strictly distinguish between semantic understanding and numerical querying, effectively eliminating the hallucination risks common in Large Language Models (LLMs) when handling critical inventory data. By coordinating the Scenario Planner and the Simulation Agent, the system performs pre execution What if simulations to quantify the trade offs between cost and service levels for different strategies. Finally, a Human in the Loop mechanism ensures high stakes decisions are executed under expert supervision, creating a safe, trustworthy, and continuously learning operational loop.

Results

Operational Outcomes The implementation of this agentic workflow aims to precisely control the balance between inventory efficiency and service availability:

Inventory Optimization: Strictly controlling the Month of Supply (MOS) for all critical SKUs within the optimal range of 0.1 to 0.13 months, significantly reducing capital stagnation.
Service Level Assurance: Maintaining a Supply Rate above 98%, ensuring uninterrupted operations at maintenance centers even under extremely lean inventory levels.

Team

National Taiwan University, Institute of Industrial Engineering (NTU IIE) & Industry Partner Joint Research Team

Prof. Jakey Blue (藍俊宏), Ph.D. Principal Investigator
- Associate Professor, Laboratory of Analytics on Knowledge Engineering (LAKE)
- Provided overall strategic guidance and validation of research methodologies, ensuring the rigorous integration of academic theory with industrial practice.
- Led the review of mathematical frameworks for supply chain optimization and coordinated resources to drive the practical implementation of the technology.
Shih Yun Lin Lead System Architect
- Graduate Researcher
- Designed and implemented the 5 Layer Multi Agent Architecture (Decision, Execution, Control, Data, Protocol layers).
- Developed core technical components: Task orchestration logic for the Manager Agent, the GraphRAG knowledge retrieval engine, and the Text to Cypher deterministic query interface.
- Constructed the Scenario Planner algorithms to quantify cost benefit trade offs in logistics decisions.
Industry Partner
- Provided deep domain knowledge of the automotive supply chain, including import lead time constraints, warehouse operation standards, and container loading logic.
- Facilitated ERP system integration and data validation, and participated in the practical testing of the Human Gate workflow.

AI-Driven Agentic Workflow for Intelligent Inventory Management