Ecommerce ops · Production

Glowe: Agent-Powered Korean Skincare Routine Recommendations Using Weaviate, Gemini, and Elysia

The problem

Most skincare recommendation systems fail to capture nuanced relationships between ingredients, user experiences, and outcomes, leaving users overwhelmed by thousands of products with no way to identify effective combinations for their specific skin.

First attempt

The common approach of embedding an entire product object into a single vector causes generic fields to dominate the vector space, diluting the importance of specific skincare effects and losing the ability to prioritize products by their actual performance.

Workflow diagram · grounded in source
1
User skin questionnaire
trigger
“The journey begins with a short, targeted questionnaire designed to understand your skin's current condition, goals, age, and any sensitivities.”
2
Product category selection
routing
“Every routine starts with a cleanser and moisturizer as a base. If it's a morning routine, we also add sunscreen. Then, based on your specific skin conditions and goals, we layer in other categories like toners, serums, exfoliators, or f…”
3
LLM review effect extraction
ai_action
“we ran every review through an LLM (Gemma 3 12B) to pull out the real effects, good and bad, that people experienced based on the review's text”
4
TF-IDF weighted effect embedding
ai_action
“we use a TF-IDF-weighted embedding based on review mentions”
5
Weaviate vector similarity search
integration
“Weaviate uses these vectors to run a similarity search against your skin needs and goals. This is done concurrently across all selected categories, resulting in a shortlist of top-matching products for each step in your routine.”
6
Gemini routine assembly
ai_action
“Gemini reviews your full profile and selects the single best product from each category, making sure they work well together, respect any sensitivities (like avoiding retinol), and avoid ingredient conflicts”
7
Agentic chat via Elysia
ai_action
“Once logged into Glowe, you get access to an intelligent, agentic chat interface alongside your personalized skincare routine. You can ask it questions, get tailored recommendations, tweak your product lineup, or find safe alternatives, …”
Reported outcome

Glowe delivers a complete morning and evening skincare routine personalized down to the ingredient level, alongside an agentic chat interface for ongoing guidance tailored to the user's skin type, concerns, and existing products.

Reported metrics
curated Korean skincare products in datasetaround 1,440
User reviews in datasetover 94,500
Reported stack
WeaviateGemini 2.5 FlashGemma 3 12BElysiaWeaviate Query AgentWeaviate Agents
Source
https://weaviate.io/blog/glowe-app
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Glowe delivers a complete morning and evening skincare routine personalized down to the ingredient level, alongside an agentic chat interface for ongoing guidance tailored to the user's skin type, concerns, and existi…

What tools did this team use?

Weaviate, Gemini 2.5 Flash, Gemma 3 12B, Elysia, Weaviate Query Agent, Weaviate Agents.

What results were reported?

curated Korean skincare products in dataset: around 1,440; User reviews in dataset: over 94,500 (source-reported, not independently verified).

What failed first in this deployment?

The common approach of embedding an entire product object into a single vector causes generic fields to dominate the vector space, diluting the importance of specific skincare effects and losing the ability to priorit…

How is this ecommerce ops AI workflow structured?

User skin questionnaire → Product category selection → LLM review effect extraction → TF-IDF weighted effect embedding → Weaviate vector similarity search → Gemini routine assembly → Agentic chat via Elysia.

Related ecommerce ops cases