culture atlas

What it does

Point at what catches your eye: an outfit, a room, a meal, or a song playing nearby. culture atlas reads its style (aesthetic, era, materials, palette and details), grounds that reading in Qloo's taste graph, and builds a living, floating feed of real brands, films, music and places connected to it.

Why it needs Qloo

An LLM can describe a photo, but it can't tell you what people who love that thing also love. Without Qloo this would be a caption generator.

• Search, with exact-name acceptance rules, identifies real brands, works, places and artists.
• Tags map style traits into the taste graph.
• Insights with explainability build "more like this" and cross-category picks across Wear, Play, Watch and Eat, each with a reason that names what connected it.
• Taste onboarding seeds the first feed from your picks, and it keeps growing from what you save and capture.

Nothing is invented

Every suggestion traces to a Qloo response. Uncertain readings are worded "resembles" and don't drive recommendations by themselves. Artists and films with no owned image get a typographic poster, never a generated portrait. An empty answer stays empty.

For agents

An MCP endpoint gives agents the same library, taste and capture tools, under scopes you grant and can revoke, so an existing agent gains cultural grounding.

How we built it

A TypeScript monorepo: Next.js for web, REST and MCP; background workers for analysis and media; PostgreSQL; native iOS (SwiftUI) and Android (Compose) apps. Vision and text models run through the Concentrate gateway. It runs on-prem with Docker Compose. Song recognition uses Apple's ShazamKit on iOS, with Qloo confirming the artist.

Challenges

Accuracy was the hard part. Models happily over-claim, so we built guardrails, a held-out evaluation set with independent judges, and strict rules for what may reach Qloo. Our photo models haven't yet passed that bar, so readings are labeled Experimental in the app rather than presented as fact. Suggestion relevance is still improving.

What's next

Better suggestion relevance, song recognition beyond iOS, and hardware testing of the Android and Meta glasses capture paths.

Built With

• docker
• ffmpeg
• kotlin
• mcp
• next.js
• postgresql
• qloo
• swift
• swiftui
• typescript