“Code twice.”
把大语言模型真正落地到业务场景：微调训练、智能体（Agent）系统、多模态大模型，以及具备长期记忆与持久化能力的 AI 应用。 Applying LLMs to real-world products: fine-tuning, agentic systems, multimodal models, and persistent long-term memory.
- 大模型结合应用场景的微调训练（SFT/对齐/评测与数据工程） Fine-tuning for practical scenarios (SFT/alignment/evaluation and data ops)
- 智能体（Agent）：工具调用、规划与决策、长期记忆与回放 Agents: tool-use, planning/decision-making, long-term memory and replay
- 多模态大模型：文本 × 语音/图像/视频 的理解与生成 Multimodal LLMs: joint understanding and generation across text/audio/vision
- 记忆持久化：让智能体在长期交互中保持一致性与“成长” Memory persistence: consistent, evolving agents across long interactions
成就 / Achievements
- ACM 铜奖 · ACM Bronze Medal
- NOIP 省二 · NOIP Provincial Second Prize
座右铭 / Motto
- “Code twice.”
- 面向业务落地的 LLM 微调与快速迭代流水线 Production-grade LLM fine-tuning and rapid iteration pipelines
- 具备长期记忆与上下文压缩能力的智能体框架 Agent frameworks with long-term memory and context compression
- 多模态理解与生成在真实链路中的集成（VLM/语音链路） Practical multimodal integration (VLM, speech) in real workflows
- 数据与评测闭环：数据策划、对齐训练、自动化评测与回归监控 Data+eval loop: curation, alignment, automated evals, regression tracking
- 长期记忆：向量检索 + 结构化记忆 + 事件时间线 Long-term memory: vector recall + structured memory + event timelines
- 多工具与多阶段规划：可解释、可回溯的复杂任务执行 Multi-tool, multi-stage planning: interpretable and traceable execution
- 端到端体验：稳定性、延迟与成本的工程权衡 End-to-end UX: reliability, latency, and cost trade-offs
- 安全与对齐：提示工程、防幻觉策略与评测基准 Safety and alignment: prompt engineering, dehallucination, and benchmarks
- 即将发布，敬请期待 Coming soon — stay tuned!
- LLM 应用工程：Agent、RAG、上下文管理、长期记忆 LLM app engineering: agents, RAG, context mgmt, long-term memory
- 多模态：文本 × 图像/语音 的联合理解与推理 Multimodal reasoning across text, image, and audio
- 数据工程：对齐数据策划、评测基准设计、回归监控 Data ops: alignment datasets, benchmarks, regression monitoring
- 系统化落地：可观测性、可扩展性、成本优化 Productionization: observability, scalability, cost optimization
- 目前不公开 · Not public for now
热衷将算法能力转化为可复用的工程资产（工具库、模板、评测框架）。欢迎就数据合成、对齐训练、Agent 框架与长期记忆策略等主题交流合作。 I enjoy turning algorithms into reusable engineering assets (libs, templates, eval frameworks). Open to collaboration on data synthesis, alignment, agents, and long-term memory strategies.