projects

Stock-Meme-Hunter is a hybrid retrieval and reranking system for uncovering hype-driven trading narratives in the A-share market.

• Proposed a dual inverted-index framework that combines pinyin-based and concept-based retrieval to capture non-semantic associations such as homophones, metaphors, and implicit market signals.
• Designed a four-stage funnel pipeline: LLM query understanding, multi-path recall, coarse filtering, and LLM reranking.
• Retrieved the top 5 most relevant stocks from a candidate pool while balancing recall quality and reasoning cost.
• Used LLM-based reranking to generate interpretable hype logic chains, making the system useful for event-driven monitoring and speculative narrative analysis.