NoDelaySCM

NoDelaySCM: An AI Agent for Supply Chain Resilience

• Description: For a global manufacturer producing complex electromechanical products like industrial lifts and smart panels, sourcing hundreds of components across a multi-tier, global factory network meant that a single delayed part could halt an entire production line. Their existing ERP systems often flagged these critical supplier delays too late, forcing costly manual interventions. To solve this, I architected and led the implementation of NoDelaySCM, an AI-powered mitigation agent that acts as a proactive co-pilot for the supply chain team. The system's intelligence is twofold: first, an XGBoost model, trained on historical SAP data and supplier performance metrics, predicts which active purchase orders are at high risk of delay. When a risk is detected, a specialized GPT-4 agent, fine-tuned for supply chain logistics using LangChain, determines the optimal mitigation strategy by analyzing part criticality, buffer stock, and alternate vendor availability.

• Key Features & Innovations:

  • Multi-Factor Delay Prediction: Trained an ML model on historic PO data, supplier performance, lead time variance, quality incidents, and factory utilization. For each active PO, classified into Green / Amber / Red zones with estimated days of slippage.

  • Context-Aware LLM Agent for Mitigation: A GPT-4-based agent (fine-tuned for supply chain context) determines appropriate action based on part criticality, buffer stock, and alternate vendor availability. Generates recommended actions: expedite, reroute from alternate supplier, escalate, or inform buffer planner.

  • Automated Workflow Execution via Cordys: Delay-risk signal (from ML) + mitigation action (from LLM) trigger workflows: Rerouting from alternate vendor (if one exists),  Triggering expediting requests with supplier via supplier portal, Alerting factory planner if no immediate solution is available and Informing project scheduling engine to re-sequence production tasks

  • Feedback Loop & Continuous Learning: Actual delivery outcomes and human overrides fed back into model/LLM retraining cycles. Accuracy improved from 73% to 87% over 6 months

• Technologies Used:

  • ML & LLM Components: Delay Prediction Model: XGBoost, trained on SAP-derived features + historical supplier behavior, LLM Mitigation Agent: GPT-4 via LangChain, with tools to access vendor maps, part criticality tables, and factory buffers, Vector DB: ChromaDB used to search alternate vendor SOPs and regulatory restrictions

  • Workflow Engine: Cordys: BPMN workflows triggered by prediction + LLM output,  n8n: Handling supplier portal messaging, alerts, and approvals, Airflow: Daily scoring jobs, weekly retraining pipelines

  • Integration & Infra

  • PostgreSQL for feature store + historical outcomes

  • Redis pub/sub for trigger coordination

  • Dockerized ML and LLM agents

  • OpenTelemetry for traceability + audit trails

• Outcome: 

• 29% reduction in unplanned production halts due to delayed parts

• 22% improvement in on-time vendor deliveries via proactive escalations

• Reused the same AI workflow engine for procurement clause monitoring and KYC document ingestion

• Recognized as a “best-practice AI blueprint” by global SCM team