Projects - Maxwell J. Yin

Product System

CineSeek

Production-style semantic retrieval system with LLM agent orchestration for natural-language movie search.

CineSeek is a production-style retrieval system combining dense embeddings, FAISS ANN indexing, and an LLM-based agent for query rewriting, reranking, and result explanation.

The agent is integrated into the retrieval pipeline to improve robustness for complex natural-language queries, enabling better handling of ambiguity, intent refinement, and long-tail search scenarios under latency constraints.

Semantic Search FAISS FastAPI LLM Agent Docker

Live Demo Source Code

Research Project

CineSeek-Adapters

Lightweight retrieval adaptation project for improving a strong frozen sentence-transformer baseline under a realistic small-data setting.

Compares linear adapters, residual MLP adapters, and concat-based item fusion under full-catalog contrastive training, with evaluation across recall, MRR, NDCG, parameter count, and retrieval latency.

The main finding is that preserving title and overview as structured item-side signals can outperform deeper nonlinear adapters, suggesting that representation structure matters more than extra capacity in this setting.

PyTorch Retrieval Contrastive Learning Ablation Study Embedding Adaptation

Source Code

Multimodal System

CineSeek-MM

CLIP-based multimodal retrieval system for text, poster, and hybrid movie search.

Implements a CLIP-based dual-encoder pipeline using Vision Transformer (ViT) image features and text embeddings, with FAISS indexing and offline evaluation.

The project analyzes modality fusion strategies, retrieval behavior, and quality/latency tradeoffs across text-only, image-only, and hybrid queries in a CLIP-style embedding space.

Multimodal Retrieval CLIP PyTorch Evaluation Image Search

Source Code