Inventory Operations (InvOps): Bringing CI/CD Principles to Inventory Management with AI-Driven Continuous Optimization

Abstract

This thesis introduces and validates INVOPS (Inventory Operations), a novel, human- centric framework designed to bridge the critical gap between data-driven algo- rithmic optimization and strategic human oversight in supply chain management. Moving beyond the established ”agentic stakeholder” paradigm of Multi-Agent Rein- forcement Learning (MARL), which operates in closed-loop simulations, INVOPS es- tablishes a Human-Stakeholder-in-the-Loop (HSITL) architecture. This is achieved through a Hybrid Kalman-RL-LLM Agent, where a Large Language Model (LLM) serves as a real-time interface to translate natural language input from actual hu- man planners into structured feedback and policy constraints for a Reinforcement Learning (RL) agent. The core innovation of INVOPS is the application of Continuous Integration and Continuous Deployment (CI/CD) principles to physical operations, creating an au- tomated, self-adapting pipeline for supply chain decision-making. This pipeline inte- grates a multi-model forecasting engine combining a SARIMAX model for context- aware demand prediction and a Kalman Filter for dynamic supplier reliability esti- mation to transform operational data into adaptive inventory policies. The frame- work shifts the paradigm from static, siloed optimization to proactive, event-driven adaptation. Empirical validation shows that INVOPS outperforms traditional deterministic ap- proaches by delivering more accurate and uncertainty-aware forecasts through a SARIMAX-Kalman hybrid, while CI/CD automation enables rapid operational adap- tation. By incorporating direct human feedback via the LLM layer, the framework aligns algorithmic optimization with strategic intent and evolving business con- straints. Overall, INVOPS provides a foundational model for resilient and intelligent supply chains that integrate human and artificial intelligence.