TasteEngine — The Cursor for Cultural Intelligence

LLM-powered cultural intelligence engine that predicts global consumer tastes in real time—by combining OpenAI’s reasoning with Qloo’s taste graph and live web insights.

Comment

Inspiration

LLMs are trained on static datasets, blind to real-time cultural shifts. So when Pepsi drops a Kendall Jenner ad and the world revolts, or Netflix flops in a new market, these models can’t warn you—it’s like using a 2020 map for 2025 terrain. TasteEngine was born to close that gap, merging OpenAI’s reasoning power with Qloo’s real-time cultural intelligence.

What it does

TasteEngine is a multi-agent LLM orchestration platform that taps into Qloo’s API and other web intelligence sources to provide live cultural analysis. It can:

  • Predict how different demographics will perceive a brand or campaign.
  • Surface real-time trends by region, age, or interest.
  • Compare cultural preferences across markets—e.g., Gen Z in Brazil vs Japan.
  • Generate visuals that match local taste using GPT-4 Vision.
  • Ingest proprietary data, perform scraping, and synthesize insights across domains.

The system executes 22 AI tools, coordinates 5 microservices, and streams its thought process live through Server-Sent Events (SSE).

How we built it

TasteEngine is built on a modular NestJS microservices architecture:

  • ai-service: LLM orchestration with OpenAI function calling.
  • qloo-service: Integrates 13 Qloo endpoints for cultural data.
  • scraper-service: Traditional scraping + vector-based semantic search.
  • scraper-v2-service: Firecrawl’s FIRE-1 agent for screenshot & visual analysis.
  • api-gateway: Manages routing, load balancing, and health checks.

All services are containerized via Docker. We also built a powerful CLI for dev testing and a real-time frontend UI powered by streaming AI outputs.

Challenges we ran into

  • Managing token limits across 22 tools while maintaining reasoning coherence was brutal—we had to engineer custom token truncation and JSON reducers.
  • Qloo’s data is powerful but complex; translating vague user queries into precise API calls required deep function tuning.
  • Streaming with SSE while juggling async scraping and image generation introduced race conditions and retry nightmares.
  • Coordinating multiple AI agents without them stepping on each other’s toes required a full multi-agent state manager.

Accomplishments that we're proud of

  • Pulled off the most comprehensive integration of Qloo’s API ever built.
  • Created a real-time LLM orchestration engine that performs complex cultural analyses in under 10 seconds.
  • Designed a CLI that lets developers test 30+ cultural tools interactively.
  • Engineered true cross-domain intelligence—connecting real-time trends, images, e-commerce data, and global cultural graphs.

What we learned

  • Cultural data isn’t just a “nice-to-have”—it’s the missing layer for any serious global-facing AI application.
  • Tool calling is far more powerful when treated as an orchestration graph, not a single turn.
  • Real-time intelligence beats stale embeddings every time—especially when working with dynamic markets like fashion, entertainment, and consumer tech.
  • Privacy-first data doesn’t mean you compromise insight—you just need smarter signals (thanks Qloo).

What's next for TasteEngine

We’re commercializing this. The plan:

  • Launch as an enterprise SaaS targeting media and CPG brands.
  • Offer API access to marketing platforms for white-labeled cultural analysis.
  • Expand to real-time influencer matching and campaign localization tools.
  • Scale infrastructure from hackathon-grade to AWS-grade, with cost optimization and region-specific deployments.

The market research industry is $76B—and 60% of that money is wasted on cultural misfires. TasteEngine is built to fix that.

Built With

Share this project:

Updates