Goal-oriented Recommendation System

Context

Most recommenders try to guess what someone wants from clicks and history — and often bury people in options that miss the mark. I built GoRS to work the other way around: the user states a goal in plain language — say, building a wall-mounted shelf — and the system writes a clear step-by-step plan, then recommends exactly the right tool for each step. It is a neuro-symbolic design. A knowledge graph, built from a real product catalogue and enriched so every item carries its purpose, key specifications, and explicitly linked alternatives, supplies precise and factual product data; a large language model does the reasoning, turning the goal into steps and matching the best-fitting product to each one. The user stays in control throughout — refining the plan in natural language, swapping any recommendation for an alternative with the differences shown side by side, or browsing the catalogue directly. Against the usual search-and-shop approach, people reached a complete, correct plan and basket far faster and with noticeably lower mental effort, were more satisfied with both the plan and the products, and most said they would buy the recommendations as they stood. By building a service around the user's goal instead of guessing it, GoRS sidesteps the cold-start and information-overload problems that limit conventional recommenders — and the same approach carries over to any setting where a goal breaks down into steps, from learning paths to treatment plans to supplier sourcing.