Internships

Contributed to the post-training of Kimi coding models across data construction, reinforcement learning, experimentation, evaluation, and model tuning, with a focus on complex software engineering and long-horizon agent tasks.

• Data synthesis: Designed synthetic training data for long-tail code tasks, multi-turn PRD/SOP scenarios, and code instruction following.
• RL training: Extended internal RL training infrastructure and ran training, tuning, and evaluation experiments.

• Reinforcement Learning
• Agentic Coding
• Data Synthesis

Worked on an agentic LLM training pipeline spanning synthetic trajectory generation, data validation, and training and evaluation environment engineering.

• Trajectory synthesis: Designed task-synthesis workflows for multi-step tool use and decision-making trajectories.
• Data validation: Built a filtering pipeline for sample cleaning, trajectory deduplication, and rule-based validation to support versioned dataset iterations.
• Environment engineering: Deployed and validated agent training and evaluation environments, including dependency management, execution debugging, and regression checks.

• Agentic Training
• Tool Use
• Data Curation

Built a government-service question-answering system using Qwen and retrieval-augmented generation, with support for multi-turn dialogue, contextual understanding, and traceable answers.

• 12345 question answering: Integrated Qwen with a RAG pipeline for the public-service hotline knowledge domain.

• Qwen
• RAG
• Multi-turn Dialogue

Reimplemented the ACTOR and ROMP 3D human reconstruction projects in MindSpore, covering core model components, training, and evaluation.

• Model reimplementation: Delivered MindSpore implementations of two established 3D human reconstruction methods.
• Open-source contribution: The resulting projects were made available through MindSpore's official community.

• MindSpore
• 3D Human Reconstruction
• Model Reimplementation