Available Aug 2026 · HK–Shenzhen 2026年8月到岗 · 港深双城

Turning data into business decisions with modern ML. 用现代机器学习，把数据转化为商业决策。

MSc AI & Business Analytics candidate at Lingnan University (HK), with an audit & finance background and hands-on experience shipping end-to-end ML systems — from FinBERT fine-tuning to reinforcement learning for routing. 岭南大学人工智能与商业分析硕士在读，具备审计与金融背景，擅长端到端机器学习系统落地 —— 从 FinBERT 微调到基于强化学习的路径优化。

Python PyTorch NLP / LLMs Reinforcement Learning PySpark Financial Analytics Trilingual 🇭🇰 🇨🇳 🇬🇧
View Projects → 查看项目 → Get in Touch 联系我

About Me关于我

Auditor-turned-ML-engineer, trilingual, built for the HK–Shenzhen corridor. 从审计师到机器学习工程师，三语沟通，专注港深创新走廊。

I'm an MSc AI & Business Analytics candidate at Lingnan University Hong Kong, graduating August 2026. Before moving into AI, I spent several years in auditing and finance at Shenzhen-based firms — that background shapes how I approach ML: I care as much about the business question and the failure mode as the model architecture.

My coursework has been systems-heavy and end-to-end: I've fine-tuned FinBERT across 32 hyperparameter configurations for financial sentiment, built a Neural Architecture Search + Deep Q-Network pipeline for vehicle routing with time windows, and shipped a PySpark text-classification system benchmarking 30 classifier-embedding pairs. I hold NVIDIA certifications in AI Infrastructure and Generative AI LLMs, and run compute on Google Colab Pro+ with A100 GPU access.

I commute daily between Shenzhen and Hong Kong, speak Cantonese, Mandarin, and English (IELTS 7.5), and my long-term goal is to build ML products that serve the HK–Shenzhen tech corridor.

我是岭南大学人工智能与商业分析硕士生，预计 2026 年 8 月毕业。在转入 AI 领域之前，我在深圳多家企业从事审计与财务工作数年 —— 这段背景决定了我对机器学习的思考方式：我既关心模型架构，也关心业务问题本身和模型的失败模式。

我的研究生课程强调端到端的系统能力：在金融情感分析中，针对 FinBERT 进行了 32 组超参数网格搜索；为带时间窗的车辆路径问题（VRPTW）构建了神经架构搜索 + 深度 Q 网络的训练框架；并基于 PySpark 完成了 30 组（5 分类器 × 6 嵌入）文本分类系统的评测。持有 NVIDIA AI 基础设施与生成式 AI 大模型两项认证，使用 Google Colab Pro+ 的 A100 GPU 进行训练。

我每天往返深港两地，熟练使用粤语、普通话、英语（雅思 7.5），长期目标是为港深科技走廊打造实用的机器学习产品。

Focus Areas专注方向

  • Applied NLP & LLMs — fine-tuning, benchmarking, and domain adaptation (finance, roleplay, Cantonese ASR)
  • Reinforcement learning for optimization — DQN, action masking, VRPTW, game environments
  • Big-data ML pipelines — PySpark, distributed training, systematic experimentation
  • Business framing — translating ambiguous business questions into measurable ML objectives
  • 应用型 NLP 与大模型 — 微调、基准评测、领域适应（金融、角色扮演、粤语 ASR）
  • 面向优化问题的强化学习 — DQN、动作掩码、VRPTW、游戏环境
  • 大数据机器学习流水线 — PySpark、分布式训练、系统化实验设计
  • 业务建模能力 — 将模糊的业务问题翻译为可度量的机器学习目标

Education教育背景

  • MSc AI & Business Analytics人工智能与商业分析硕士
    Lingnan University, Hong Kong 香港岭南大学     Expected Aug 2026预计 2026.08
  • Prior: Auditing & Finance此前：审计与金融
    Shenzhen-based firms 深圳多家企业

Certifications认证

  • NVIDIA — AI Infrastructure
  • NVIDIA — Generative AI & LLMs
  • IELTS 7.5

Languages语言

  • Cantonese粤语Native母语
  • Mandarin普通话Native母语
  • EnglishIELTS 7.5

Skills技能

Technical depth · business framing · production tooling. 技术深度 · 业务建模 · 工程工具链。

</>

Technical技术能力

Python SQL PyTorch Transformers PySpark XGBoost NLP Deep RL LLM fine-tuning Statistics
📊

Business商业能力

Problem framing问题定义 Metrics design指标设计 A/B testingA/B 实验 Storytelling数据叙事 Financial analysis财务分析 Auditing审计 Reporting汇报
🛠

Tools & Stack工具与栈

pandas scikit-learn Hugging Face PostgreSQL MongoDB Git / GitHub Colab Pro+ (A100) Selenium / Playwright KNIME Tableau

Projects项目

Four featured projects across NLP/finance, time-series forecasting, reinforcement learning, and combinatorial optimization. 四个精选项目，覆盖 NLP/金融、时序预测、强化学习与组合优化。

FinBERT Fine-Tuning for Financial SentimentFinBERT 金融情感分析微调

CDS525 · Mar–Apr 2026
PyTorchTransformersFinBERTGrid Search
Problem问题Financial texts (earnings reports, analyst notes) use domain-specific language that generic sentiment models misclassify — a real cost for any analyst or investor relying on automated signals.金融文本（财报、研报）高度依赖领域词汇，通用情感模型误判率高 —— 对依赖自动化信号的分析师或投资者而言代价不菲。
Data数据Financial PhraseBank (~4.8k sentences, 3-class sentiment). Handled moderate class imbalance with weighted cross-entropy and focal loss.Financial PhraseBank 数据集（约 4800 句，3 类情感）。使用加权交叉熵与 Focal Loss 处理类别不平衡。
Approach方法Full factorial grid search: 2 loss functions × 4 batch sizes × 4 learning rates = 32 training runs, all tracked in a unified notebook with reproducible seeds.完整析因网格搜索：2 种损失函数 × 4 组 batch size × 4 组学习率 = 32 次训练，统一 notebook 可复现。
Outcome成果Best macro F1: 89.92% (Weighted CE, BS=32, LR=2e-5). Delivered unified training notebook, ~2,664-word Word report, and GitHub README.最佳 Macro F1：89.92%（加权交叉熵, BS=32, LR=2e-5）。交付统一训练 notebook、约 2664 字 Word 报告与 GitHub README。
My Role我的贡献Group project — I designed the experiment grid, implemented the unified training loop, ran all 32 configurations, and authored the analysis.小组项目 —— 我负责实验设计、统一训练循环实现、执行全部 32 组实验并撰写分析报告。
GitHub →

NAS-EA + DQN for Vehicle Routing with Time Windows基于 NAS-EA + DQN 的带时间窗车辆路径优化

CDS526 · Feb–Apr 2026
Deep RLPyTorchNeural Architecture SearchEvolutionary Algorithms
Problem问题Last-mile delivery with time-window constraints is NP-hard. Classical solvers are exact but slow; naive RL agents pick infeasible actions and train unstably.带时间窗的最后一公里配送是 NP 难问题。传统求解器精确但慢；朴素强化学习智能体频繁选择不可行动作，训练不稳定。
Data数据Solomon VRPTW benchmark instances. Customer locations, demands, time windows, and service durations.Solomon VRPTW 基准数据集：客户坐标、需求、时间窗与服务时长。
Approach方法Combined Neural Architecture Search via evolutionary algorithm with a Dueling DQN agent. Implemented action masking so infeasible actions get Q=−∞, and diagnosed depot-reset and time-window feasibility bugs across v6→v7→v8.结合进化算法驱动的神经架构搜索与 Dueling DQN。引入动作掩码使不可行动作 Q=−∞，并在 v6→v7→v8 过程中修复了 depot 重置与时间窗可行性缺陷。
Outcome成果v7 achieved mean distance 731 with 18% CV, vs v6's 948 / 77% CV — a 23% improvement in solution quality and 4× improvement in stability across 5 seeds.v7 取得平均路径长度 731、变异系数 18%，相较 v6（948 / 77%）解的质量提升 23%，稳定性提升 4 倍（5 个随机种子）。
My Role我的贡献Individual research project. Designed NAS search space, implemented masking, ran multi-seed experiments, and wrote the report.个人研究项目。负责架构搜索空间设计、掩码机制实现、多种子实验与报告撰写。
GitHub →

European Weather Forecasting: XGBoost vs BiLSTM欧洲天气预测：XGBoost 对比 BiLSTM

CDS524 · 2025–2026
XGBoostBiLSTM + AttentionTime SeriesFeature Engineering
Problem问题Temperature forecasting benchmarks usually pit tree models against deep learning with standard lag features. But can domain-informed features — in this case Chinese cosmological calendar features (24 Solar Terms, Wuxing, Heavenly Stems) — actually shift the balance? And do the two model families respond to them the same way?温度预测的常见评测是用标准滞后特征对比树模型和深度学习。但领域启发式特征 —— 这里指中国历法体系（二十四节气、五行、天干） —— 是否真能改变两者的表现平衡？两类模型对这类特征的响应是否一致？
Data数据European daily weather records with temperature, humidity, pressure, and wind features. Aligned to the Chinese solar calendar for feature engineering.欧洲日度气象数据：温度、湿度、气压、风速等。与中国农历节气体系对齐后构造特征。
Approach方法Clean 2×2 factorial design: {XGBoost, BiLSTM+Attention} × {standard lag/rolling features, Chinese cosmological features}. Statistical significance tested with paired t-tests across multiple seeds.清晰的 2×2 析因设计：{XGBoost, BiLSTM+Attention} × {标准滞后/滚动特征, 中国历法特征}。多随机种子配对 t 检验评估显著性。
Outcome成果Non-obvious finding: Chinese calendar features significantly helped XGBoost without lag features (p=0.0005) but harmed LSTM performance — suggesting the two architectures treat cyclical priors very differently. Delivered full report, GitHub README, and a 20-minute narrated PPT.非直观结论：中国历法特征对不使用滞后特征的 XGBoost 显著提升（p=0.0005），但损害 LSTM 表现 —— 表明两类架构对周期性先验的利用方式截然不同。交付完整报告、GitHub README 与 20 分钟讲稿的 PPT。
My Role我的贡献Group project. I designed the factorial experiment, engineered the Chinese-calendar feature set, built both pipelines, and ran the statistical analysis.小组项目。我负责析因实验设计、中国历法特征工程、两类流水线实现以及统计显著性分析。
GitHub →

Space Defender: Dueling DQN + Prioritized Experience ReplaySpace Defender：Dueling DQN + 优先经验回放

CDS524 · 2025
Deep RLDueling DQNPERCurriculum LearningReward Shaping
Problem问题A standard RL benchmark game — but a good testbed for a real-world issue: agents that exploit reward loopholes (e.g., "corner-hiding" to maximize survival time without engaging) rather than learning the intended behavior. How do you get stable, high-scoring, and honest play?这是一个标准强化学习基准游戏，同时是现实问题的良好载体：智能体往往会钻奖励漏洞（例如"躲角落"来延长存活时间而不参与对抗），而非学习真正期望的行为。如何训练出稳定、高分且行为符合预期的策略？
Data数据Self-generated from a custom Space Defender game environment — state, action, reward, next-state tuples sampled during rollouts across 8 training iterations.自定义 Space Defender 游戏环境自产生数据 —— 8 个训练迭代中采样的 (状态, 动作, 奖励, 下一状态) 经验元组。
Approach方法Dueling DQN with Prioritized Experience Replay, iterated v1→v8. Diagnosed reward-hacking (corner-hiding), redesigned the reward function, and introduced curriculum learning in v8 to progressively harden the enemy spawn pattern.Dueling DQN 结合优先经验回放（PER），迭代 v1→v8。诊断出奖励黑客行为（躲角落），重新设计奖励函数，并在 v8 中引入课程学习，逐步加大敌方生成难度。
Outcome成果Reached mean score ~1,960 (target 3,500). More importantly, the iteration log documents how each architectural and reward-shaping change moved the needle — a realistic record of RL debugging, not just a final number.最终平均得分约 1960（目标 3500）。更重要的是迭代日志完整记录了每次架构与奖励调整的效果 —— 真实呈现强化学习调试过程，而不只是给出一个终值。
My Role我的贡献Individual project. Full ownership: environment, agent architecture, PER implementation, reward design, and 8-version iteration.个人项目。独立完成：环境搭建、智能体架构、PER 实现、奖励设计与 8 版本迭代。
GitHub →
Additional projects: CDS527 PySpark Starbucks text classification benchmark (5 classifiers × 6 embeddings, Tuned BERT+LR reached 71.6% accuracy); Cantonese ASR benchmark (Whisper fine-tuning on Common Voice 17.0 for a CMHK R&D internship application); CDS529 LLM roleplay benchmarking (BFI-2 personality + Empath voice fidelity across 5 LLMs × 100 characters); CIFAR-10 CNN dissertation. 其他项目：CDS527 PySpark 星巴克文本分类基准（5 分类器 × 6 嵌入，调参后 BERT+LR 准确率 71.6%）；粤语 ASR 基准（基于 Common Voice 17.0 的 Whisper 微调，为应聘中国移动香港 R&D 实习准备）；CDS529 大模型角色扮演基准（BFI-2 人格 + Empath 语料保真度，5 个大模型 × 100 角色）；CIFAR-10 CNN 毕业论文。

Resume简历

Quick snapshot — PDF version available below. 简要快照 —— 完整 PDF 版本请见下方下载。

Education教育

Lingnan University, Hong Kong香港岭南大学 Expected Aug 2026预计 2026.08

MSc in Artificial Intelligence & Business Analytics人工智能与商业分析硕士

Experience工作经历

Auditor / Finance Associate审计师 / 财务专员 Shenzhen-based firms深圳多家企业

Audit engagements, financial analysis, and reporting. Built the business-framing mindset I now bring to ML work.负责审计项目、财务分析与汇报工作。所培养的业务思维已融入当前的机器学习工作。

Selected Projects精选项目

FinBERT Grid Search — 32-config fine-tuning, best macro F1 89.92%32 组超参数微调，最佳 Macro F1 89.92%

NAS-EA + DQN VRPTW — 23% solution-quality gain, 4× stability improvement解的质量提升 23%，稳定性提升 4 倍

Weather Forecasting 2×2 Factorial — XGBoost vs BiLSTM with Chinese calendar features (p=0.0005)XGBoost 对比 BiLSTM，中国历法特征显著性 p=0.0005

Space Defender Dueling DQN + PER — 8-version iteration, reward-hacking diagnosis, curriculum learning8 版本迭代，奖励黑客诊断，课程学习

Certifications & Languages认证与语言

NVIDIA AI Infrastructure · NVIDIA Generative AI & LLMs · IELTS 7.5

Cantonese (native) · Mandarin (native) · English (professional)粤语（母语）· 普通话（母语）· 英语（专业工作级）

Download PDF Resume ↓ 下载 PDF 简历 ↓ LinkedIn →

Contact联系方式

Open to ML/Data Science roles across Hong Kong and Shenzhen. I reply within 24 hours. 欢迎港深两地的机器学习 / 数据科学岗位机会，24 小时内回复。

@
Email邮箱 neofastio@gmail.com
GH
GitHub zhoubojian-stevenchow
in
LinkedIn bojian-zhou
☎
Phone (HK)电话（香港） +852 9102 1799