Pin-Zu Li
I'm Pin-Zu Li, an LLM engineer based in Taipei with a passion for building practical AI systems and RAG workflows.
About
I’m Pin-Zu Li, an LLM engineer based in Taipei with a passion for building practical AI systems and RAG workflows. At Taiwan AI Labs, I co-designed a multimodal RAG architecture, built an extensible evaluation module, and developed text-modality benchmarks, achieving a 7% performance improvement on the zh-TW benchmark. I also trained a dense retriever to support English queries retrieving Chinese documents and led the development of an agentic RAG with context engineering to assess query complexity and adapt retrieval and reasoning.
Earlier at the TAIDE project (National Institutes of Applied Research), I created a large-scale data processing pipeline to curate Taiwanese culture content, contributed to TAIDE’s performance toward GPT-3.5-level Chinese, and led post-training optimization for the release version. I proposed chat vectors to transfer instruction-following capabilities from English LLMs to Traditional Chinese LLMs and published practical experiences on localizing LLMs for Taiwan.
- 1 year's experience
- Joined October 2025
Languages
- ChineseFluent
- EnglishAdvanced
Industry experience
- Software & Internet
- Media & Entertainment
- Professional Services
More experts in Taipei, Taiwan
Work experience
Data Collection and Processing: built a large-scale text processing pipeline to extract, filter, and curate high-quality web content reflecting Taiwanese culture. Academic TAIDE Development: achieved performance comparable to GPT-3.5 in Chinese at the time. Post-Training Optimization: developed the post-training process for the release version of TAIDE. Chat Vector: proposed chat vectors to transfer instruction-following capabilities from English LLMs to Traditional Chinese LLMs without additional training.
RAG Development: Co-designed a multimodal RAG architecture with an extensible evaluation module and text-modality benchmarks, achieving a 7% improvement on the zh-TW benchmark; trained a dense retriever for English queries retrieving Chinese documents. Agentic RAG: built a context-engineered RAG to assess query complexity and adapt retrieval/reasoning, boosting in-house benchmark by 2 points; developed an RL-based small model to optimize the Agentic RAG process, delivering 4x faster performance with comparable benchmarks.
Education
Hire Pin-Zu Li today
Post a job and invite Pin-Zu, or contact them directly. No hidden commission.