BOLT Cultural Intelligence Control Plane

Inspiration

A cultural recommendation system needs more than generic popularity: it needs an actual taste evidence source and a human-controlled boundary before a recommendation becomes an action. I built BOLT TasteGuard, a small Qloo-powered prototype through Unfire, to make that boundary visible.

What it does

The user enters an audience, a cultural anchor (for example, the artist Rosalía), an objective, and a recommendation category. The Node.js server looks up the anchor via the official Qloo Search API and obtains its true Qloo entity ID. That ID is used in a Qloo Insights API request for up to eight recommendations. The browser renders the returned entities as text, displays the query context, and requires an explicit human approval step before any hypothetical follow-up. Approval has no side effects; it does not book, pay, post, or contact anyone.

How I built it

• Server: Node.js 20+, official Qloo REST endpoints /search and /v2/insights.
• Secret handling: QLOO_API_KEY is used server-side only, never stored in client code or the repository.
• UI: plain JavaScript, safe text rendering of external recommendations, explicit approval state.
• Reliability: missing keys, invalid options and API failures are surfaced instead of generating invented recommendations.
• Open source: MIT-licensed source code and eight Node.js automated tests.

Code, setup and test instructions: https://github.com/ondmindmanagement-hub/bolt-tasteguard

Verification (October 8, 2026)

I tested the official API with the event-issued key: Qloo Search returned HTTP 200, and Qloo Insights returned HTTP 200. An end-to-end call resolved Rosalía as the taste anchor and retrieved eight actual place recommendations. Separately, 8/8 automated tests passed. The public code was updated and pushed. These are local integration results, not a production deployment, commercial pilot or evaluation of recommendation quality.

Important limitations

The form's city and audience fields are context only: they are not geographic or demographic filters in the Insights request. The place recommendations can therefore be worldwide and must not be represented as venues in the entered city. The first Qloo Search match is used; ambiguous anchor names require future disambiguation work. The workflow is deterministic API orchestration with a human approval UI; it is not a fully autonomous agent or proof of market traction.

What I learned

Integrating a real taste graph is substantially more useful than giving an LLM a prompt and treating its guesses as evidence. The harder product lesson was to be precise about what the data actually supports, surface missing filters, and avoid any implicit authority to act.

Built by Omar Baró / Unfire in Spain; Unfire is a founder-led initiative in a pre-incorporation stage. The technical prototype was created and reviewed with AI coding assistance and validated with real official API responses.

Built With

• ai-agents
• api
• automation
• cultural-intelligence
• governance
• javascript
• qloo
• recommendation-systems