Seeking Position: Data Analyst / Business Intelligence Intern
Turning data into stories with Python, SQL, and machine learning to solve real business problems
Turning data into stories with Python, SQL, and machine learning to solve real business problems
I am a Master's student in Artificial Intelligence and Business Analytics, expected to graduate in June 2026. Passionate about extracting business value from large-scale data, I combine strong statistical foundations with programming skills to drive decisions through A/B testing and machine learning models.
đ Education Background
Lingnan University(Hong Kong)
Master of Science in Artificial Intelligence and Business Analytics (2025 â 2026 expected) | GPA 3.6/4.0
Dalian Minzu University
Bachelor of Artificial Intelligence (2021 â 2025) | First-Class Scholarship
Problem: Accurately analyze emotional polarity in Twitter user comments to understand public opinion, especially handling simple vocabulary, emojis, and complex symbols.
Data: Twitter user comments dataset containing varied text (simple words, emojis, and complex symbols).
Approach: Built a neural network with embedding layer, Dropout, bidirectional LSTM layer, and fully connected layer; trained using Adam optimizer; performed literature review, model reproduction, and performance improvements.
Outcome/Impact: Achieved 98% accuracy on simple-vocabulary comments, 91% on emoji/complex-symbol comments, and 88% overall test-set accuracy. The model showed strong adaptability across comment types and provided empirical support for LSTM-based sentiment analysis.
Your contribution: Team member â innovated and optimized the model after reproducing baseline, organized literature review, visualized results and training charts, created presentation PPT, proofread/improved project report, wrote main paper content, and refined format/structure for accuracy and logical rigor.
View Acceptance Letter (EEIC 2024)Problem: The complex multi-sided food delivery platform (Meituan Waimai) required reverse engineering of its database to support user membership, payment processing, and review systems while ensuring data integrity and query efficiency.
Data: Conceptual data derived from core app functionalities (user profiles, addresses, payments, transactions, reviews, restaurants, riders).
Approach: Identified entities, attributes, and primary keys; designed ER diagrams for three modules (membership, payment, review); created relational schema and normalized tables to 3NF; designed denormalized queries for user behavior analysis.
Outcome/Impact: Delivered a coherent, normalized relational database schema that supports high-frequency transactions, real-time status updates, and efficient analytics dashboards; enabled targeted marketing through user segmentation and membership value analysis.
Your contribution: Team member â contributed to entity/relationship modeling, ER diagram creation, relational schema design, normalization process, and user behavior data analysis design.
Problem: Create an interactive maze navigation game where an agent must learn optimal paths in a stochastic environment with moving traps, collectible coins, and random maze layouts.
Data: Dynamically generated 10Ã10 mazes (perimeter + random internal walls) with real-time state representation (agent, goal, traps, coins, walls).
Approach: Implemented Deep Q-Network (DQN) using PyTorch (3-layer NN, replay buffer, target network, Îĩ-greedy); custom dense reward function with distance shaping; Pygame for rendering, animations, UI, and manual/AI modes.
Outcome/Impact: Agent converged to positive rewards (+303.4 max in 4 steps), win rate improved from <5% to ~25%; created an educational tool demonstrating reinforcement learning in dynamic, uncertain environments with real-time visualization.
Your contribution: Individual work (environment design, DQN implementation, reward engineering, UI, training pipeline, and documentation).
âļī¸ View YouTube Demo đ View GitHubProblem: Build and optimize deep learning models for accurate classification on the challenging CIFAR-100 dataset (100 fine-grained classes, small 32Ã32 images, limited samples per class).
Data: CIFAR-100 dataset (60,000 color images: 50,000 train, 10,000 test; split 80/20 train/validation + official test).
Approach: Preprocessing (normalization + augmentation: random crop, horizontal flip); baselines (linear model, MLP with ablation on hidden size/activation/dropout/weight decay); advanced residual-style CNN with BatchNorm, dropout, and residual connections; controlled experiments on loss, LR, and batch size.
Outcome/Impact: Best MLP configuration achieved 36.20% test accuracy; comprehensive ablation studies and learning curves demonstrated hyperparameter impact; provided clear insights into transitioning from simple to advanced architectures for real-world computer vision tasks.
Your contribution: Implemented training pipelines, conducted all ablation experiments, generated learning curves, accuracy tables, visualizations, and performance analysis (group project).
đ View GitHub