RuiTian

NLP AI Engineer Intern

• Built graph-augmented RAG pipelines with Milvus, achieving 100% file-level recall and reducing latency from 64.3s to 36.8s on financial documents.
• Developed entropy-regularized zero-shot GRPO for sample-efficient LLM training, matching baseline performance with 75% fewer rollouts on math datasets.
• Improved financial information extraction, raising temporal extraction accuracy from 69.0% to 93.0% and stock code accuracy from 54.5% to 97.9%.

Pharmolix

AI Developer Intern

• Built a Milvus–FastAPI hybrid retrieval system with query rewriting and chunk-level fusion for document QA.
• Developed a PDF-grounded annotation tool enabling accurate chunk-to-document alignment.
• Constructed a LightRAG-based knowledge graph for context-aware retrieval across large corpora.