Lookalike

1. Dual-Engine Matching Mechanism

• Precision Layer: Used ArcFace/InsightFace for high-dimensional facial embedding extraction and Cosine Similarity for structural accuracy.

• Semantic Layer: Leveraged LMM (e.g., Gemini Vision) to decode high-level "vibe" (aesthetic, makeup, expression), providing reasoning beyond simple pixel matching.

2. Automated Database Scaling

• Dynamic Sourcing: Built a pipeline to scrape and filter rising influencers, ensuring the database stays culturally relevant.

• LMM Auto-Tagging: Used Multimodal Models to automatically index profiles with style/persona tags, stored in Faiss/Milvus for sub-second vector retrieval.

similar project:  Presearch's Drag & Drop Creator Discovery