Qice Sun

About

AI/LLM engineer with production experience building a multimodal RAG agent for Walmart driver support, combining hybrid retrieval (dense+BM25) and fine-tuned Llama 3 served via vLLM on Azure AKS to reach sub-second latency. Drove measurable impact (25% fewer escalations, 60% lower token costs, 33% lower storage costs) and also built Kafka-based microservices that cut batch runtime from 2 hours to 15 minutes and reduced DB load by 80%.

Key Strengths

• Built and deployed production multimodal RAG agent for driver support with spatial/visual context
• Reduced recurring ticket escalations by 25% via self-improving knowledge distillation loop
• Achieved sub-second Llama 3 inference latency by serving via vLLM on Azure AKS
• Cut storage costs by 33% by offloading raw binaries to Azure Blob and optimizing ETL
• Reduced API token costs by 60% using tiered memory caching and prompt caching
• Designed metric-driven deployment gates using RAGAS (context precision/recall/faithfulness) to control hallucinations
• Architected Kafka-based microservices platform; reduced batch runtime from 2 hours to 15 minutes and cut DB load by 80%
• Effective cross-functional delivery with operations stakeholders using KPI-based communication

Experience

Education

Skills

Languages