The video explores the intersection of e-commerce and modern agentic technology, focusing on how the latest AI advancements can be integrated with traditional machine learning algorithms to enhance e-commerce analytics and customer engagement. E-commerce has historically been a fertile ground for machine learning due to its vast amounts of structured data and the potential for significant financial rewards. However, large language models (LLMs), which power much of the recent generative AI wave, struggle with processing large volumes of structured data directly. The solution presented involves using proven predictive algorithms as tools for LLM-based agents, allowing the AI to leverage both traditional data science methods and generative capabilities effectively.
Central to this approach is Kumo’s Relational Foundation Model (Kumo RFM), which combines generative language models with graph neural networks (GNNs). GNNs excel at understanding complex relationships within large, interconnected datasets, making them ideal for e-commerce data. The video demonstrates a fully functional chat application where an agent uses Kumo RFM alongside traditional data querying tools to analyze an H&M e-commerce dataset. This agent can identify valuable customers, predict their likelihood of making purchases, and suggest personalized marketing strategies, such as targeted email campaigns with product recommendations based on predicted future purchases.
The tutorial then dives into the technical setup, showing how to authenticate and initialize Kumo RFM, load and prepare the e-commerce dataset, and convert it into pandas data frames for easier manipulation. The speaker explains how to define semantic types and relationships between tables to build a graph structure that Kumo RFM can use for predictive queries. Using Kumo’s predictive query language (PQL), the model can forecast customer behavior, product demand, and churn probabilities. The video highlights the speed and accuracy of these predictions compared to what would be possible with LLMs alone.
Next, the video covers building an AI agent using a lightweight framework called Graph AI. This agent integrates two main tools: one for querying pandas data frames and another for querying Kumo RFM. The agent uses OpenAI’s GPT-4.1 model to generate queries and interpret results, dynamically deciding which tool to use based on the user’s questions. The agent’s architecture is designed as a simple graph with nodes representing the language model, tools, and start/end points, enabling it to handle complex multi-step interactions efficiently. The speaker also discusses practical considerations like limiting the number of steps per interaction to avoid excessive querying and managing long customer IDs for better performance.
Finally, the video demonstrates how to deploy the entire system as a Dockerized application with a front-end chat interface and back-end API, making it easy to set up and use. The speaker emphasizes the broad applicability of this approach, not only for internal analytics but also for customer-facing agents that can provide personalized shopping assistance. By combining Kumo RFM’s predictive power with generative AI’s natural language capabilities, businesses can create highly responsive, data-driven e-commerce agents that improve marketing effectiveness and customer experience. The video concludes by inviting feedback and encouraging viewers to explore the provided code repositories to build their own agentic e-commerce solutions.
