Inspiration
Independent music venues, cultural shops and small brands know their audience, but translating that taste into merchandise is difficult. A plausible product list is not enough: a merchandiser needs to understand the cultural fit, check what is in stock, preserve margin and approve the final collection.
CultMerch AI connects those decisions in one traceable workflow.
What it does
Describe an audience and optionally add named artists, films or brands with explicit priorities. CultMerch resolves cultural references, queries Qloo, and assembles a collection from a 50-product synthetic catalog.
The workflow:
- Interpret the brief and choose cultural anchors and Qloo tool requests.
- Exclude products with zero stock or a gross margin below the chosen threshold.
- Resolve catalog brands and request Qloo cultural tags and targeted brand affinities.
- Compare two collections using the same brief interpretation, catalog and commercial constraints, with Qloo signals added only to branch B.
- Inspect each product's evidence, request revisions and approve every item before JSON or CSV export.
A revision locks downloads, the JSON preview and clipboard copying until the collection is fully approved again.
How we built it and what Qloo contributes
The application uses Next.js, React, TypeScript, Tailwind CSS and Zod. Its bounded agentic workflow orchestrates Qloo searches, affinity requests, catalog filtering, ranking and human review.
Qloo /search resolves audience anchors and the brands represented in relevant catalog products. /v2/insights supplies cultural tags and evaluates targeted brands against the audience anchor. Affinity scores and reference priorities affect product ranking. The interface retains source entities, timestamps, request outcomes and explicit product matches.
The bridge from a cultural signal to a SKU is inspectable: exact tags, local editorial translations and declared catalog-brand associations are distinguished. A brand association does not mean that brand produced or endorsed the synthetic product.
The hosted version uses disclosed local rules for interpretation and copy. An optional Gemini adapter exists in the source but is not enabled in this deployment. We do not claim a live Gemini reasoning step.
What we verified
A fresh hosted analysis on 30 October 2026 for fans of John Coltrane, Japanese raw denim, analog audio and handmade ceramics completed 12 logical Qloo operations and collected 53 signals. All recorded provider calls succeeded, and all four selected products had explicit Qloo evidence.
Adding Qloo changed two of four selected products. The synthetic basket changed from EUR365 to EUR498, with calculated gross margin changing from EUR200.50 to EUR261. This EUR60.50 difference is theoretical: no sales, conversion uplift or independent improvement in taste quality was measured. The baseline is a controlled comparison, not a benchmark against every LLM.
The video replays that saved real result locally, explicitly labeled Cached Qloo, with its original timestamps. The public application can run new analyses. Tests also checked approval persistence, revision locking and JSON/CSV exports. The final local suite recorded 94 passes, zero failures and one deliberately skipped live test; offline tests use simulations rather than being counted as new Qloo calls.
Challenges and what we learned
Broad cultural signals need an explicit matching layer before they can influence a feasible collection. Unmatched signals remain available in provenance, while the highlighted insight summary contains only evidence used by selected products. Missing affinity scores do not receive an invented bonus.
We also learned that approval must cover every export path, including preview and clipboard copying. Those paths now share the approval gate, and copied JSON matches its displayed preview.
A higher internal alignment score is not proof of purchase intent. The human reviewer needs to see the evidence and retain the ability to disagree.
What's next
Merchant-owned catalogs, shared durable storage, broader brief interpretation, commerce connectors and prospective commercial evaluation. These are future work.
Try it
- Application: https://cultmerch-ai.onrender.com
- Public MIT source and setup instructions: https://github.com/Robe092/cultmerch-ai
- English demonstration: https://youtu.be/blvLDmySAlQ
No login is required. The free host may need about a minute to wake up. Start with a brief such as "Fans of John Coltrane, Japanese raw denim, analog audio and handmade ceramics", click Run analysis, then inspect Collections, Catalog and Approvals. All products are synthetic. Saved analysis links can expire because storage is temporary and designed for a single server. Testing instructions are in TESTING_EN.md in the repository.
Built With
- next.js
- node.js
- qloo
- react
- render
- tailwindcss
- typescript
- zod
Log in or sign up for Devpost to join the conversation.