Constructing the advertisment platform and system.

Built a shopping AI agent for E-commerce recommendations using Large Language Models (LLM) and OpenAI function-call.

• Optimized embedding-based Retrieval-Augmented Generation (RAG) with Milvus vector database by integrating 1) synthetic merchandise review with expert transcripts and 2) query rewrite via Chain-of-Thought (CoT), boosting F1 score by 37%.

• Constructed an evaluation system to score workflow traces based on context relevance (RAG eval.) and answer accuracy (LLM eval.), leading to continuous prompt refinement and a 36% increase in Click-Through Rate (CTR).

• Developed a security monitoring platform for prompt injection detection, enabling real-time intercepts for data exfiltration.

• Led the infrastructure security team in constructing a comprehensive cloud-native security system with Scrum agile mode.

• Derived a zero trust architecture from 0 to 1 across the entire enterprise, spanning multiple clouds and on-premises networks.

• Enabled hierarchical fault-tolerance with Raft algorithm using gRPC, enabling end-to-end error detection and auto-recovery.

• Implemented a unified authentication module in the L7 load balancer (reverse proxy), maintaining ≤ 10ms additional latency scaling 10x hot-pluggable rules (coordinated via Redis) across peak traffic of 100k QPS.

• Deployed services in AWS EC2 and serverless Lambda with Terraform, achieving auto-scaling for 1k-100k fluctuating.

• Protected 1k+ services by identifying and mitigating 100+ threats monthly, achieving a NPS of 80 among 10k+ users.

IP 技术预研与技术标准化