The Agent That Knows Taste

Inspiration

Agents handle logistics beautifully and taste terribly. Ask one to set up a client dinner and it books the nearest mid-range restaurant: correct, and instantly forgettable. The missing input is not logistics, it is affinity.

What it does

AOS_engine is a working agentic copilot layer: it classifies intent, plans tool calls, streams its work, and holds a confirmation boundary before anything consequential executes. For this hackathon we give it taste. A get_affinity(context, audience) tool joins the registry, backed by Qloo's affinity engine, so the agent can pull cultural-affinity signals mid-plan: the audience's tastes across cuisine, music, venues and categories. Venue shortlists, gifts and outreach drafts are then filtered by affinity, not just distance and price.

How we built it

The provider pattern already in the codebase (LLM_PROVIDER flips between OpenAI, Gemini, OpenRouter key pools and a mock provider with no code changes) extends to Qloo: one more provider, one more set of tools. Per-client data isolation and key-pool rotation apply to the affinity calls, and the confirmation boundary still governs every consequential action: advise with taste, act with permission, audit everything.

Demo flow

"Set up the dinner with the Paliwal family next week. Make it land." The plan decomposes, get_affinity runs, the shortlist is taste-filtered, and the booking waits for one human yes.

Challenges

Keeping the agent honest: the Qloo integration is the build of this hackathon, while everything about AOS_engine's shipped features comes straight from its README.

Built With

• fastapi
• openrouter
• python
• qloo