Jianxiang Gao

Engineered an LLM-based document audit platform processing ~1% of U.S. patient medical records, building a generalized clinical document processing pipeline with 99.99%+ extraction accuracy and reducing manual implementation time by 90%. Designed marker-based learning to reduce parsing cost/latency while improving accuracy. Built agentic onboarding and support workflows using Next.js and Temporal (90% faster ticket resolution, 75% lower engineering overhead). Developed an internal coding agent saving 80%+ engineering time and reducing customer support cycle from 3 weeks to 1 day. Architected AI evaluation pipelines with prompt version management and agentic workflow observability, improving overall system performance by 30%.

Selected as 1 of 10 global teams for Amazon Alexa Prize TaskBot Challenge 2; built an LLM-powered multimodal voice-and-vision task assistant using Llama and CLIP to guide users through multi-step tasks on Echo devices. Designed an agent backend with microservices on Amazon ASK API and EC2, using GPU Docker, caching, ECS load balancing, and auto-scaling to support a 50M+ user base and cut operational latency by 50% under heavy load. Developed a graph-based conversational framework and a hybrid safety alignment system, reducing conversation friction and increasing user response rate by 30%. Analyzed production logs to debug and fix backend service issues, communicating findings directly with Amazon engineers.

Led end-to-end development of three production systems—a video-processing microservice, a financial AI chatbot, and an admin panel—each shipped independently on AWS. Built a video query and ranking system using vector databases on a scalable AWS pipeline (S3, Textract, CloudFront, Lightsail, MySQL), improving search relevance/performance by 55%. Engineered a financial chatbot with LangGraph, GPT models, and Ray, improving retention and engagement by 30%. Designed and built an admin panel with React/TypeScript/Ant Design Pro and Flask, reducing administrative cost by 68%.