Developer fine-tunes a 350M LLM to run as a shopping agent entirely in the browser

A developer has fine-tuned LiquidAI's LFM2.5 models (230M and 350M parameters) into a functional front-end shopping agent that runs entirely in the browser, requiring no server, API key, or cloud infrastructure. The model was trained on behavioral patterns rather than domain-specific facts, allowing it to power multiple different storefronts — such as a coffee shop or grocery store — using a single set of weights without retraining. It uses a fixed set of eight tools and retrieval-augmented generation to handle tasks like adding items to a cart, resolving references, and answering policy questions based on runtime-injected context. Training involved roughly 30 million tokens of synthetic data generated from around 18 interaction recipes, with a teacher model (Qwen3 30B) writing natural-language content and code handling structured tool calls. The project highlights on-device AI benefits including user privacy, offline functionality, and zero inference costs, with a live demo available on GitHub Pages.