AI Shopping Assistant (NDIS Marketplace)

This POC explored an AI assistant for an NDIS-focused Magento marketplace. That made the problem sharper than a normal retail chatbot. The catalog was complex, the intent could be highly specific, and the cost of hallucination was not trivial. A wrong answer would not just be awkward. It could mislead users in a context where accuracy mattered.

The architecture leaned toward grounded retrieval instead of model memory. The assistant pulled from the Magento REST API, catalog data, and vector search so responses stayed tied to real products instead of drifting into generic generated advice. That choice follows a pattern I trust: when truth already lives in changing operational systems, retrieval plus live checks is usually safer than asking the model to "know" too much.

The POC also made it clear that "assistant" was hiding multiple jobs. Search, comparison, and action are related, but they are not the same workflow. A single prompt can fake all three in a demo, but a production system usually needs more explicit structure if it is going to stay safe and useful.

The main output from the POC was not a vanity demo. It was a set of better evaluation questions. Does the assistant narrow options well? Does it stay inside catalog truth? Does it improve assist-to-cart rate? How often does it resolve the query correctly without inventing confidence it has not earned?

The trade-off was giving up the fantasy of a simple chatbot. Grounded assistants take more engineering. They need indexing strategy, retrieval tuning, fallback handling, and clear boundaries around what the model can infer. They may also feel less "creative," but in commerce that is usually a good thing.

This was a disciplined prototype. It reduced uncertainty around the part that actually matters: whether the assistant could be trusted enough to help people buy the right thing without making the platform harder to trust.