Inspiration

Korean K-pop fans have a word for copying exactly what your idol wore or where they went: "Son Min-soo". Fans hunt for the bag an idol carried at the airport and the café they visited. A friend who helps with this is a dukme, a fan-mate.

We analyzed 2,519 messages over 66 days from two real fan open-chat rooms. In one room alone, 156 fans asked about 5 "what is this item?" questions a day. The median wait for a first answer was 80 minutes (over 11 hours in the other room), and 32 questions were never answered. And when fans do find the item, it's often ₩1,600,000 or sold out.

Generic LLMs can't fix this. They guess idol facts, and they guess taste.

What it does

FanMate combines two kinds of knowledge:

Facts from fans: what idols actually wore (brand, product, price, source) and where they actually went, from our fan-verified dataset (Son Min-soo Wiki + reCree). It never invents idol facts.
Taste from Qloo: when the original is out of budget, Qloo's taste graph finds brands with the same taste profile. After planning a pilgrimage route, Qloo adds nearby places that fit your taste. Love NCT WISH? Qloo suggests artists.

It remembers your biases, budget, sizes, wishlist, purchases and visited spots, and replies in English, Korean, Japanese or Chinese. Every answer shows Fan-verified cards and Qloo taste match cards separately, so fans always know what's a fact and what's a recommendation.

Testing instructions

The live demo runs on a public sample of our data: BTS, IVE, aespa, RIIZE and NCT WISH (74 items; 75 spots in Seoul and Busan). Qloo recommendations work for any brand, area or artist. Try:

"Show me what Karina from aespa wore"
"Wonbin's knit vest is over my budget ($50). Find me something with the same vibe"
"Plan a RIIZE pilgrimage in Seongsu and add places that fit my taste"
"I bought the vest!" or "I went to Seongsu", then ask again. It remembers (see "My Fan Log").
Ask in Korean or Japanese and the whole UI switches.

How we built it

Agent framework: Croche (tool use, long-term memory, character expressions) on a Node.js/TypeScript server with a web chat UI and Leaflet maps
9 tools: 8 over our own data and per-user state, plus tasteMatch, which calls Qloo /search (resolve brands, spots and groups), /v2/tags (vibe → tags) and /v2/insights (brand, location-filtered place, and artist recommendations), seeded with what the fan told us they love
Deployed on Render (always-on server, per-visitor memory isolation, rate limits)
The repo includes a public sample of our data; the full dataset (600+ items, 500+ spots, 30+ groups) is private

Challenges we ran into

Joining two worlds: our data knows what an idol wore, while Qloo knows taste. We map idol items and pilgrimage spots onto Qloo entities and keep facts and recommendations visibly separate.
Qloo returns brands, not products with prices, so the agent is honest that it's recommending brands to explore.
Fans write in four languages and fandom slang, so the agent keeps the reply language stable and resolves nicknames to real members.

What we learned

Grounding matters twice. LLMs guess both idol facts and taste: real fan data fixed the first, and Qloo fixed the second.

What's next

We start with K-pop, the most intense "copy my idol" culture in the world. But "what did they wear, where did they go, what fits me?" is a question fans of every celebrity, athlete and influencer ask. Sonminsoo Lab already runs an in-house automated pipeline that collects and verifies "Son Min-soo" data (items and places tied to a public figure) and feeds Son Min-soo Wiki. FanMate is the consumer layer on top: as the pipeline expands from K-pop to global stars and creators, fans of any star can use the same agent, with Qloo turning every verified fact into recommendations that fit their own taste and budget.