Dong Wei

I build production-minded LLM agent systems that diagnose and measurably improve agent behavior. My work spans orchestration, observability, evaluation, retrieval, and streaming AI application delivery.

For FinLab-X, I’m building a financial research agent across news search, SEC 10-K RAG, and Text-to-SQL data layers using Python, FastAPI, LangGraph, LlamaIndex, Qdrant, and DuckDB. I reduced median time-to-first-visible-token from about 36 seconds to about 4 seconds, improved ticker-precision@10 from 0.62 to 1.00, and built a Braintrust evaluation framework around a 30-question golden dataset.

I also shipped Jin-T, a bilingual Chinese-English speech-to-text application. I used tracing to analyze output failures and designed a code-plus-LLM fallback architecture, while removing transcription length limits with chunked parallel processing and per-chunk streaming.

Previously, I developed Form Recognizer Studio and open-source labeling experiences at Microsoft Taiwan AI R&D Center for Azure AI Document Intelligence. I also speak and write about agent observability, evaluation, RAG retrieval failures, and hybrid AI architectures.