chen xie

Coding Agent Eval Contractor

Snorkel AI

Contributed to Terminus Bench, an open-source coding agent evaluation benchmark with sandboxed Docker environments for safely executing and testing AI-generated code. Designed new agent-eval benchmarks and integrated a LoRA fine-tuned DeepSeek V4 Flash model into the benchmark pipeline to improve assessment performance on coding agent tasks.

PythonDockerLoRAsdata preprocessingTokenization

Part-time Contractor

Snorkel AI

Designed and implemented coding agent evaluation benchmarks to expand task coverage and improve assessment reliability. Fine-tuned DeepSeek V4 Flash with LoRA SFT using ms-swift and integrated the model into the benchmark pipeline to measure performance gains.

PythonDockerLoRAsDeepSeek V4 FlashSwiftdata preprocessingTokenizationBenchmark Integration

Senior LLM Engineer

Built a multi-agent, multi-turn report generation system using RAG with short- and long-term memory. Implemented SFT plus DPO-style RL optimization to improve dialogue quality and increased key engagement metrics (transition rate and CTR), and developed an iterative Search-R1-inspired RL framework within a modular multi-LLM search pipeline.

PythonPyTorchRAGMulti Agent SystemsQwenSFTDPOReinforcement Learning

Software Engineer (Azure ML)

Microsoft

Implemented fine-tuning pipelines on Azure ML Studio for multiple base models (including Llama3, GPT-4o, and Mistral) and managed training data via Azure Blob Storage. Built multilingual training pipelines, generated SBOM artifacts for data and model training/deployment steps, and evaluated fine-tuned models for Microsoft Teams meeting summaries.

PythonPyTorchRAGAzure Blob StorageModel Fine TuningSBOMsMultilingual NLP Pipelines

Software Engineer (ML Platform)

Designed and implemented a cloud-native ML training platform deployed across two Kubernetes clusters, including user authentication and REST APIs for training-data management. Improved performance using Redis caching (reduced data retrieval time by 60%) and built task-queue architecture with Redis and gRPC, plus monitoring/alerting in Prometheus and Grafana that reduced false alerts from 46.6% to

GoCKubernetesMicroservicesgRPCRedisREST APIsPrometheusGrafanaVue.Js