Search and AI
Vibe2Value
An influencer search engine that ranks creators by semantic fit and estimated commercial value.
- Role
- Full-stack AI engineer
- Timeline
- RoCathon project
Overview
Vibe2Value helps brands search for creators using the language they naturally use when describing a campaign. It turns a subjective brief into ranked results by combining semantic similarity with estimated sales potential.
The challenge
Keyword search is a poor fit for creative briefs. A brand may describe a creator as calm, minimal, technical, or energetic without using the exact words stored in a creator profile. The ranking also needs to consider business value, not only language similarity.
The approach
I built a local embedding and vector-search pipeline so the product could understand meaning without depending on paid external AI APIs. Express and Node.js provide the service layer, PostgreSQL with pgvector stores embeddings, and Zod validates data at the application boundary.
Architecture
Semantic representation
Transformers.js generates local embeddings for creator profiles and search briefs.
Vector search
PostgreSQL and pgvector retrieve creators with similar meaning, not only matching keywords.
Ranking
Semantic relevance is combined with estimated sales potential for business-oriented results.
API
Node.js and Express expose the search workflow to the frontend.
Validation
Zod keeps request and response data explicit at the system boundary.
Outcomes
- Made natural-language creator discovery possible without paid external AI APIs.
- Combined semantic relevance with a commercial ranking signal.
- Built a complete search product suitable for a hackathon demonstration and further iteration.
What I learned
- Local models can be a practical choice when cost, privacy, or deployment control matters.
- Search ranking should reflect the user's real decision, not just the similarity score.