Experience
-
ByteDance
- Led the development of self-improving Agent infrastructure, building feedback loops for experience extraction, evaluation, and continuous optimization.
- Designed a multi-scope Agent memory architecture supporting Shared, Session, Task, and Run-level isolation and reuse through sandbox-based runtime integration.
- Built a sandbox execution infrastructure using containers and MicroVMs for secure code execution and browser automation, covering resource provisioning, scheduling, workspace persistence, and lifecycle management.
-
Tencent
- Led the development of an MCP-based sandbox infrastructure, integrating Agent runtimes with Docker for container lifecycle management, stateful Python execution (IPython/Jupyter Kernel), interactive shells, and browser automation via VNC.
- Re-architected a RAG pipeline with hybrid retrieval, NER-based intent analysis, reranking, and streaming responses, improving modularity and extensibility.
- Worked on GPU inference infrastructure, optimizing throughput and resource utilization through batching, quantization, KV-cache management, and deployment improvements.
-
Meituan · Dianping
- Built a GraphRAG-based content moderation system integrating multimodal retrieval with CLIP embeddings and evaluation pipelines, improving reviewer productivity by 8% and moderation accuracy by ~2 percentage points.
- Designed a DAG-based asynchronous processing system with fault tolerance, graceful degradation, and distributed tracing, increasing throughput by 15% and reducing average latency by 40%.
- Contributed to a request aggregation and deduplication system using Redis and Kafka, reducing server count from 55 to 27, redundant requests by 72%, and peak database write pressure by 45–84%.