Building Agents for Ecommerce (kumo.ai)

The piece showcases a practical blueprint for building e‑commerce agents that combine generative LLMs with scalable, relational predictive models — specifically Kumo’s Relational Foundation Model (KumoRFM), which fuses LLMs’ reasoning with Graph Neural Networks’ (GNN) strength on structured, relational data. Rather than forcing an LLM to ingest billion‑row tables, the agent pattern gives the model a set of tools (e.g., KumoRFM for predictions, dataframe query tools) so the LLM orchestrates analytics and natural-language interaction while specialized components handle heavy lifting. That hybrid approach addresses a key gap: generative AI excels on unstructured text, but e‑commerce predictive tasks still require structured-data primitives like GNNs and ANNS. Technically, the article walks through loading an H&M sample (customers, articles, transactions) into KumoRFM via rfm.LocalTable and rfm.LocalGraph, inferring metadata, linking foreign keys, and initializing model.predict calls (e.g., PREDICT SUM(..., 30 days) or COUNT(...)=0) that return probabilistic outputs (TARGET_PRED, False_PROB/True_PROB). It then shows agent wiring: an LLM core (OpenAI gpt‑4.1‑mini by default) plus GraphAI as a lightweight execution framework and custom tools (a QueryDataframes tool that execs pandas code in a sandboxed namespace). The result is an on‑the‑fly forecasting agent that preserves confidence scores, avoids heavy pretraining, and makes it straightforward to combine natural language, ad‑hoc analytics, and large‑scale relational prediction — a pattern likely to accelerate sophisticated personalization, demand forecasting, and automated decisioning in e‑commerce.