AI Engineer & AI Agent Developer

AI 是新时代的
电力
AI is the new
electricity.

— Andrew Ng, 2016

嗨,我是 David。我专注于 AI 应用落地,将业务流程、知识与决策逐步 Agent 化,让 AI 从回答问题的工具,变成能够理解任务、调用工具并完成交付的执行系统。 Hi, I'm David. I work on getting AI into production: turning workflows, knowledge and decisions into agents step by step, so that AI stops being a tool that answers questions and becomes a system that understands a task, calls the tools it needs, and delivers.

我相信,AI 的价值不在参数、Demo 或简单替代人工,而在于真正改变生产方式、释放生产力,并放大人的判断力、创造力与行动力。因此,我选择保持 Builder 的本能:贴近问题,快速交付,并将每一个工作流、每一次交互和每一处细节,打磨成真正值得使用的体验。 I believe the value of AI lies not in parameters, demos or simply replacing people, but in genuinely changing how work gets done, unlocking productivity, and amplifying human judgement, creativity and initiative. So I keep a builder's instincts: stay close to the problem, ship fast, and polish every workflow, every interaction and every detail into something genuinely worth using.

01

经历Background

2019.09 — 2023.07Sep 2019 — Jul 2023

利物浦大学University of Liverpool

理学学士 · 数学与统计BSc · Mathematics & Statistics

2023.10 — 2025.10Oct 2023 — Oct 2025

帝国理工学院Imperial College London

理学硕士 · 计算机科学与技术MSc · Health Data Analytics & Machine Learning

2025.10 — 2026.11Oct 2025 — Nov 2026

中国科学院北京基因组研究所(国家生物信息中心)Beijing Institute of Genomics, CAS (China National Center for Bioinformation)

助理工程师Assistant Engineer

2026.12 起From Dec 2026

深度原理Deep Principle

AI Agent 开发工程师AI Agent Development Engineer

02

项目实践Selected work

工作项目 · 中国科学院北京基因组研究所At Beijing Institute of Genomics, CAS

GSA 数字员工GSA Digital Worker

面向科研数据审编与文献取证,独立搭建支持长任务执行、跨会话记忆与权限治理的 Agent 系统。An agent system I independently built for research-data curation and literature evidence gathering, with long-running tasks, cross-session memory and access controls.

以 SOP 状态机与 Agent Loop 协同执行,接入版本化 Skills、MCP 工具与生命周期 Hooks。在 60 个中断与超时用例中,任务恢复成功率达 95%;人工审编耗时由 30–60 分钟缩短至 5–10 分钟。Combines an SOP state machine with an agent loop, versioned skills, MCP tools and lifecycle hooks. Task recovery succeeded in 95% of 60 interruption and timeout test cases, and manual curation time fell from 30–60 to 5–10 minutes.

Agent HarnessHooksSkills · MCPMemoryAccess Control
工作项目 · 中国科学院北京基因组研究所At Beijing Institute of Genomics, CAS

文献多跳检索与 Search Agent 后训练Literature Multi-hop Retrieval & Search Agent Post-training

围绕文献—数据集关系核验,结合混合检索、多跳证据验证与模型后训练,向数字员工提供可定位原文的证据包。Hybrid retrieval, multi-hop evidence verification and model post-training for literature–dataset relationship checks, returning evidence packages traceable to source passages.

构建 1,350 条逐跳证据 QA,并按论文与研究项目隔离数据。在固定 200 条领域测试题、相同工具与调用预算下,Qwen3-4B 经 LoRA SFT / GRPO 的答案正确率为 54% → 68% → 76%。Built 1,350 evidence-grounded QA examples, split by paper and research project. On a fixed 200-question domain test set with the same tools and call budget, LoRA SFT and GRPO improved Qwen3-4B answer accuracy from 54% → 68% → 76%.

Agentic RAGHybrid RetrievalQwen3-4BLoRA SFTGRPO · verl
个人项目 · 独立设计与实现Independent project · Design & implementation

EvoWork · 自进化 Agent HarnessEvoWork · An Evolving Agent Harness

从零实现 Runtime、Memory、Skills 自进化、Eval 与安全五个子系统,探索让 Agent 从任务失败中学习,并通过回归验证后集成技能。Built runtime, memory, skill evolution, evaluation and safety subsystems from scratch, exploring how agents can learn from failures and integrate skills after regression validation.

打通“轨迹分析 → 失败归因 → 技能提案 → 沙箱回归 → 确认集成”。经 6 轮迭代,固定 60 Case 回归集成功率从 52% 提升至 80%。Connects trace analysis, failure attribution, skill proposals, sandbox regression and confirmed integration. Across six iterations, success on a fixed 60-case regression suite rose from 52% to 80%.

RuntimeLayered MemoryContext EngineeringSkill EvolutionEvaluation

a GSA 数字员工:人工审编耗时。GSA Digital Worker: manual curation time.

b Search Agent 后训练:Qwen3-4B 答案正确率。Search Agent post-training: Qwen3-4B answer accuracy.

c EvoWork:回归集成功率。EvoWork: regression-suite success rate.

Fig. 1 | 项目结果。Fig. 1 | Project results. 灰为改进前,蓝为改进后。Grey: before; blue: after.
03

能力地图Skill map

Agent 运行框架Agent Harnesses

Agent LoopReActPlan-and-ExecuteHooksSkills · MCPMulti-Agent

记忆与上下文Memory & Context

Episodic / Semantic滚动摘要Rolling Summaries上下文压缩Context Compression工具结果外部化Tool-output Externalization

检索与后训练Retrieval & Post-training

Agentic RAGGraphRAGBM25 · BGE-M3RRF · RerankLoRA SFTGRPO · verl

评测与工程Evaluation & Engineering

Ragas · LLM-as-JudgeEM / F1RBACPython · FastAPILangGraphDocker · GitClaude Code · Codex