HZ
GitHub LinkedIn
Research & Engineering

Hanlin Zhao

Exploring how AI systems
learn, perceive, and act.

I’m an undergraduate researcher at CUHK, working at the intersection of machine learning, computer vision, and embodied intelligence, with co-authored work on robotic endoscopy.

GitHub LinkedIn Curriculum vitae Email

Currently

Undergraduate Researcher

Artificial Intelligence:
Systems & Technologies
The Chinese University of Hong Kong

Seeking Summer 2027 internships
Machine learning · Computer vision · Embodied AIExplore selected work
01 /

About

I study AI systems, from the experiments that help us understand them to the engineering that makes them work.

I’m pursuing a BEng in Artificial Intelligence: Systems & Technologies at The Chinese University of Hong Kong, with expected graduation in 2029. Alongside my studies, I conduct undergraduate research in the AI Robotic Perception Lab under Prof. Hongliang Ren.

My interests include machine learning, computer vision, embodied AI, vision-language-action models, self-evolving AI, and recommender systems. I enjoy both research and implementation: designing experiments, training models, and building software with reproducible evaluation.

I’m seeking AI/ML and software engineering internships for Summer 2027, and I’m interested in future research opportunities and research-based graduate study.

Beyond research: Chinese Debate Team and participation in the Kaggle CASMI26 molecular identification competition.

02 /

Research

Undergraduate research · CUHK

AI Robotic Perception Lab

Supervised by Prof. Hongliang Ren

My work in self-evolving AI includes experimental design, experimental execution, and model training within the lab team.

Papers & preprints

OpenReview profile

Other lab research

Lab research topic

EndoNav

Endoscopic navigation.

Lab research topic

EndoRinse

Robotic irrigation.

EndoNav and EndoRinse are team research topics; individual responsibilities and unpublished experimental details are not described here.

03 /

Selected projects

All repositories

Implementation, experiments, and the evidence behind the results.

01Independent project

Movie recommendation system

Explicit-feedback recommendation with five rating predictors and biased matrix factorization implemented directly in NumPy.

  • Mini-batch SGD, validation-based selection, and cold-start fallbacks.
  • Saved model artifacts, a recommendation CLI, and full-catalogue Top-10 evaluation.
Recorded result

Test RMSE 0.8578 vs. 1.0389 for the global mean on one MovieLens split. Better rating accuracy did not translate into better Top-10 ranking than the constant-score baselines.

Evaluation & limitations
  • Python
  • NumPy
  • pandas
  • pytest
View repository : movie recommendation system
02AIST1000 seminar project

Background sensitivity in CLIP

An inference-only study of how object-centric preprocessing affects CLIP predictions under paired background interventions.

  • Frozen CLIP ViT-B/32 across 370 Oxford-IIIT Pet images and 11 conditions.
  • Segmentation-based crops, matched controls, and class-stratified paired bootstrap intervals.
Recorded finding

Original-image accuracy was 83.78% for full images and 83.24% for object crops. Both paired intervals for flip-rate reductions included zero; a robustness benefit remains unresolved.

Results & experimental scope
  • Python
  • PyTorch
  • OpenCLIP
  • NumPy
View repository : background sensitivity in CLIP
03ESTR2018 course project

Exploration & exploitation

A probabilistic study of stationary Bernoulli bandits, linking concentration bounds to simulated regret.

  • Uniform random, fixed ε-greedy, and UCB1 policies implemented in NumPy.
  • Independent random streams, paired trials, regret scaling, and Hoeffding experiments.
Experimental scope

The main comparison uses 300 trials of 5,000 rounds. Saved trial data and uncertainty intervals support comparisons; finite simulations do not establish asymptotic regret rates.

Read the course report
  • Python
  • NumPy
  • Matplotlib
  • unittest
View repository : exploration and exploitation
04Developed from ESTR1002 coursework

Maze Explorer

A C11 terminal game with procedurally generated mazes, limited visibility, and shortest-path hints.

  • Seeded depth-first maze generation and breadth-first pathfinding with wrapping portals.
  • Game-state logic, energy constraints, and a native Windows console interface.
Engineering focus

Separates maze logic from terminal rendering. Includes CMake builds and logic tests covering 1,200 seeded mazes; the repository records Windows verification.

Inspect the logic tests
  • C11
  • CMake
  • Windows API
  • curses
View repository : Maze Explorer

More coursework in C, game theory, and R is available in cuhk-coursework .

04 /

Education

The Chinese University of Hong Kong

2025–2029 · Expected

BEng in Artificial Intelligence: Systems & Technologies

CGPA 3.883 / 4.000Dean’s List 2025/26

Relevant coursework Recommender Systems (STAT3009), Regression Analysis (STAT3008), Mathematics and Statistics.

Honors & scholarships

  • Dean’s List2025/26
  • Lee Quo Wei CUHK Golden Jubilee Scholarship
  • CUHK Entrance Scholarship
05 /

Curriculum vitae

A concise record of my work.

Education, research, papers, and selected engineering projects.
Updated October 2026.

Download CV PDF · One page
06 /

Contact

Let’s talk about research
and building useful AI.

I’m interested in Summer 2027 AI/ML and software engineering internships, research collaborations, and opportunities to learn from experienced teams.

zhl2104405665@gmail.com Connect on LinkedIn Find me on GitHub OpenReview