TakeStock: AI for Supply Chain Insight

Complex supply chains and growing U.S. threats endanger the Defense Industrial Base. Supply chains are complex, especially for DoD systems, whose sourcing and demand constraints are very different from commercial companies. Deciding how to deploy our limited resources requires analyzing those supply chains, but their complexity makes analysis hard and requires more supply chain analysts than are often available. Existing analysis tools are hard to use if you haven’t spent a lot of time getting to know them. Logistics experts can’t get the insight they need to make smart decisions. An easier way to extract information from supply chain data is critical. Using AI to make insight available to anyone. Ever since ChatGPT arrived near the end of 2022, large language models (LLM) have revolutionized how users work with computers. They can generate text in response to user input that mimics what a human would respond. Through the power of their statistical analysis of text, they can reproduce the human reasoning encoded in that text. However, they are fundamentally a tool for generation, not understanding, especially in problem domains that involve new and changing data. They hallucinate – they make stuff up – and sound confident doing it. Hallucinations can’t be avoided, and fine-tuning, where an LLM is further trained on specific problem domain data, only makes hallucinations worse. Retrieval Augmented Generation (RAG), where outside information is fed into the LLM along with the user’s input, reduces hallucinations, but even that doesn’t eliminate them. The LLM struggles if any RAG information is noisy or contradictory, and it’s limited by LLM’s limited context windows (like its attention span). LLMs on their own, even with RAG, can’t analyze data and provide insight. Enter TakeStock. nou Systems, Inc. (nSI) and the MIT Center for Transportation and Logistics (MIT CTL) will deliver TakeStock, software that lets users ask questions and get back insight in the form of summaries, charts, and graphs. It will query data from existing Earth 616 graph databases, forecast scenarios using advanced machine learning (ML) models, and generate scenarios and test cases from user questions. An LLM provides a natural language interface and formats and summarizes output.