Aarcade Assistant

Players and builders in the Aavegotchi ecosystem ask the same questions constantly — how the DAO is structured, what an installation costs, where to trade — and the answers are scattered across wikis, forum threads, and contract state. A static FAQ goes stale and a general-purpose chatbot has no idea what a Baazaar is.

The assistant runs entirely on Cloudflare: a Pages front end in Vue, Pages Functions for the API, Gemma through Workers AI for generation, and Vectorize for retrieval. A router classifies each message and picks a path — bundled ecosystem knowledge, retrieval over documents ingested from the wiki, a live web search, a GitHub or Basescan lookup, an image analysis, or plain conversation — and the model is grounded with whatever that path returns. Skills are bundled as documents the router can prefer for governance questions, and a later pass added image attachments with combined vision-and-search answers and pixel-art generation through the PixelLab API. It ships as an iframe embed so it can sit inside the arcade site.

Live and embeddable, running from February 2026 through a year of iteration: knowledge routing first, then retrieval, then vision and generation. The trading research scaffold that grew alongside it was extracted into Gotchi Trader. This was the first product built around a hosted model and is where the routing-and-grounding pattern used in later agent work was worked out.