Inventory optimization is heavily based on understanding material data and patterns in the data. Our expertise includes material analysis for complex supply chains, demand and lead time analytics, getting insights for wahrehouse and material networks, building algorithms to optimize millions of customer materials, developing forecasting and risk management tools, material matching algorithms using NLP and graphs.

In particular, we found that advanced techniques in material simulation and optimization (see eg. our dashboards ), probabilistic modeling are crucial in finding optimal inventory policies: please see Figure 1 below. They allow finding optimal inventory policies even for a more sophisitcated multi-echelon supply chains (see Figure 2.). It can be done using a simulate-and-optimize approach as shown in Figure 1, or Deep Reinforcement Learning as shown in Figure 3. The latter approach is more computationally intensive, however typically gives a better results: please see Figure 4.

Figure 1. Inventory simulation and optimization

Figure 2. Inventory optimization in a multichelon setting.

Figure 3. Inventory optimization using Deep Reinforcement Learning.

Figure 4. Inventory optimization using Deep Reinforcement Learning abd Bayesian Optimization.

At the same time, NLP models are very helpful to "describe" materials, and graph models (see 'Graphical representation of supply chain' on this page ) help undestand a place of a given material in a complex supply chain are extremely helpful. The last two tools combined allow us to find identical/similar materials and reduce costs related with inventory replanishment.

The latter can be done using embeddings for material descriptions and material graphs. Those embeddings being stored in a vector databases (like Pinecone, PGVector, ChromoDB, etc), which allows finding similar items quickly and effectively.