Cue — Cultural Intelligence for Creative Discovery

Inspiration

Creative discovery is about more than finding something that matches a category. A filmmaker creating a skateboarding montage, for example, needs music that captures the scene's energy, atmosphere, and style—not simply a list of songs from a particular genre.

AI can interpret a creative brief and generate plausible suggestions, but meaningful recommendations also need cultural context. We wanted to explore how cultural intelligence could become a functional part of an AI recommendation system rather than just an additional layer of explanation.

That idea inspired Cue: an application that combines AI-powered creative intent interpretation with Qloo's cultural intelligence to help users discover recommendations that better match their creative vision.

What We Built

Cue transforms natural-language creative briefs into culturally informed recommendation slates.

Users describe the mood, genre, atmosphere, style, and context they want. Cue interprets that intent and integrates Qloo's API into its recommendation pipeline to help identify and rank relevant candidates.

Cue focuses on three core capabilities:

Creative intent interpretation: Translates natural-language briefs into signals that guide recommendations.
Cultural grounding: Uses Qloo's cultural intelligence as part of candidate discovery and ranking.
Transparent recommendations: Lets users compare Qloo-enabled and Qloo-disabled results and communicates when the system must broaden a genre search.

One of Cue's distinguishing features is its Qloo ON/OFF comparison. Users can compare the culturally grounded recommendation slate with an alternative from the same run. Switching the comparison does not require another recommendation request, making the influence of cultural grounding easier to inspect.

Cue also handles incomplete genre coverage explicitly. When Qloo returns no candidates for an exact genre, the application can attempt a bounded broader-genre match when an appropriate alternative is available. It tells users when that happens instead of presenting broader results as exact matches. If a suitable scoped result cannot be found, Cue can disclose when it falls back to an unscoped ranking.

How We Built It

We built Cue as a web application with an integrated recommendation pipeline combining natural-language interpretation, Qloo's API, candidate filtering and ranking, and a comparison interface.

The pipeline begins with a creative brief. Cue interprets the user's intent and identifies signals that can guide the recommendation process. Qloo's cultural intelligence then contributes to candidate discovery and ranking.

During development, we refined the genre-filtering logic to ensure that only relevant music-genre tags restrict music candidate pools. This helps prevent unrelated cultural tags from incorrectly narrowing the recommendations.

We also implemented bounded broader-genre fallback behavior using Qloo's own tag index. Rather than inventing a genre hierarchy or silently substituting unrelated candidates, Cue looks for an appropriate broader genre when the exact match produces no results and clearly communicates the change.

The comparison interface reuses the same run's results, avoiding extra recommendation requests when users switch between Qloo-enabled and Qloo-disabled views.

Finally, we developed automated verification tests for important parts of the recommendation pipeline, including genre filtering, fallback behavior, and pipeline correctness. The application also supports AI-provider fallback when its primary provider is unavailable.

Challenges We Faced

Interpreting Cultural Tags Correctly

One significant challenge was that a cultural tag returned by an API is not automatically suitable for every recommendation category. We had to refine Cue's filtering logic so unrelated tag readings would not incorrectly scope music recommendations.

This taught us that integrating a knowledge API requires careful interpretation of its data, not simply passing returned tags directly into a recommendation pipeline.

Handling Missing Genre Matches

Some niche genre queries returned no artists from Qloo. A naive system might silently return unrelated recommendations or present a broader category as though it were an exact match.

We addressed this by implementing a bounded broader-genre fallback and making the change visible in the interface. When an appropriate broader match is unavailable, Cue can disclose when it has to fall back to an unscoped ranking.

This preserves user trust by distinguishing an exact match from a broader alternative.

Making Cultural Grounding Observable

We wanted users to see the practical difference cultural grounding makes, rather than simply asking them to trust a claim about AI and cultural intelligence.

The Qloo ON/OFF comparison allows users to inspect alternative slates from the same run without triggering another recommendation request. Building this behavior required us to think carefully about how to make the comparison meaningful and efficient.

Managing AI Provider Availability

AI providers can experience temporary outages or quota limitations. We implemented provider fallback behavior so a failure from one provider would not necessarily prevent Cue from producing recommendations. During live testing, Groq successfully handled requests when Gemini was unavailable.

What We Learned

Building Cue reinforced an important lesson: a plausible AI-generated answer is not necessarily a well-grounded recommendation.

Recommendation quality depends on how well the system interprets user intent, selects relevant candidates, uses cultural context, and handles incomplete data.

We learned to treat cultural tags as domain-specific signals that need validation, rather than assuming every returned tag is relevant. We also learned that fallback behavior is part of the product experience: when the system cannot satisfy a request exactly, it should explain what changed and preserve as much of the original intent as possible.

Another important lesson was the value of testing the entire recommendation pipeline. It is not enough to verify that an AI model returns text. We also need to examine how external data, filtering, ranking, and provider failures affect the final result.

Accomplishments

We are proud to have built a working creative recommendation experience in which Qloo's cultural intelligence contributes to the recommendation pipeline.

Our key accomplishments include:

Integrating Qloo into creative recommendation discovery.
Implementing genre-aware candidate filtering.
Building bounded broader-genre fallback behavior.
Communicating when recommendations use a broader genre or an unscoped fallback.
Implementing a Qloo ON/OFF comparison without additional recommendation requests when toggled.
Adding automated tests for recommendation filtering and fallback behavior.
Testing the live application with different creative briefs, including punk rock, alternative rock, and shoegaze scenarios.

Most importantly, we focused on building a system whose behavior can be inspected and whose limitations are visible to users.

What's Next?

We want to expand Cue's usefulness across creative workflows, including music discovery for film, video, and other media projects.

Future improvements could include better coverage for niche genres, stronger evidence that helps users understand individual recommendations, and broader evaluation of how cultural grounding affects relevance across different creative briefs.

Our goal is to make creative discovery more context-aware without sacrificing transparency.

Conclusion

Cue explores a simple idea: understanding what someone wants is only part of making a useful recommendation.

By combining AI-powered intent interpretation with Qloo's cultural intelligence, Cue turns creative briefs into recommendation slates that are more context-aware and whose behavior can be inspected.

We built Cue to demonstrate that cultural intelligence can be a functional part of an AI recommendation pipeline—and that when the system cannot find an exact match, it should say so.

Live Demo: https://cue-nine-iota.vercel.app/

Source Code: https://github.com/itzebe/cue