Nice to meet you! I love creating little gadgets and dive into the possibilities of AI.

• Energetic and enthusiastic towards learning new things.
• Have experiences with PyTorch and Tensorflow projects
• Graduating with a major in AI in Summer 2025
• 🏆 Dean's List, 2021-2022, 2022-2023 and 2023-2024 in CUHK

• 🚀 Skills:
  • AI:
    • Frameworks: Python, Tensorflow, PyTorch, HuggingFace, Keras
    • Models:
      • Reinforcement Learning: Deep Q-Learning
      • Vision: ResNet, U-Net, GANs, Diffusion
      • NLP: RNN, LSTM, BERT, LLaMa
  • Backend: Python, Java, MySQL, Docker, Flask
  • Frontend: JavaScript, CSS, HTML
  • Planning to Learn: Node.js, React

Jun 2024 - Aug 2024 | Summer Internship
Sep 2024 - Nov 2024 | Part-time

• Implemented and trained RNN-based models for sign language translation tasks
• Developed a backend pipeline for data fetching and LLM inference, integrating it with existing fine-tuning workflows using Python, Flask, and MySQL.
• Researched methods for temporal alignment on gesture sequences

Oasis: The Calendar App (Jun 2024)	SnowFight: Deep Q-Learning Agent for Third-Person Shooter Game (Dec 2022)	RegSubjer: Course Registration with Autoclicker (Jan 2022)

Mar 2023 - Present

• Independently developed a mobile event planning and notification application for Android using Java, aiming to help users track their deadlines and events
• Worked across the development lifecycle to build and maintain code
• Migrated the data storage and retrieval to the more robust RoomDatabase with extensive use of SQL queries
• Refactored codebase into distinct UI layers and business logic components

Others:

• Link: https://github.com/ash3327/OasisPlanner/tree/development
• Latest Release: Still in development phase.
• Skills: Java · RoomDatabase · SQL

Sep 2024 - Dec 2024

• Gesture detection and mapping into the 3d scene with three.js
• Beatmap generation via librosa

https://github.com/ash3327/ai_music_project

Jan 2024 - Apr 2024

https://github.com/ash3327/Peer-to-Peer-Communication-App

• Created a peer-to-peer communication app supporting audio recording, waveform display and editing, and also screen share function
• Implemented GUI for audio recording, waveform display and other functions.
• Implemented synchronization mechanism for audio and video packets sent through socket.
• Skills: Python

• A customizable RPG game with map creation tools.
• Developed a GamePlayer feature allowing users to play their maps or load other people's maps.
• Created a MapCreator tool for users to change landscape tiles, place items, monsters, and NPCs, and add multiple levels to their maps.
• Implemented a Custom Code Editor enabling users to customize map interactions, including NPC interactions through custom coding.
• Link: https://github.com/ash3327/ArchaicBitmapGame
• Skills: Java

• Created a Gym environment of a simple third-person shooter game in Python
• Implemented a Deep-Q Network with PyTorch to train agents to master the game with a variable quantity of moving objects
• Fine-tuned the model to achieve average kill streak of 7 and lengthen survival duration by 4 times, significantly better than the random baseline.

With limited information provided to the agent, the agents developed the following strategies without prior prompting:

Precise Shooting	Retreats to Corner	Constant Spinning

Exploring the generalizability of Vision Transformers (ViTs) on small datasets compared to Convolutional Neural Networks (CNNs). The project highlights:

• Scalability: ViTs underperform on datasets with very small sizes, while CNNs are more robust.
• Efficiency: ViTs are computationally less efficient than CNNs for models with the same accuracy.
• Models used: ResNet, ViT, DeiT, and T2T-ViT.

Also touched training U-Net over the Cityscape dataset.

	Accuracy Comparison	Computational Efficiency
Classification
	Training IoU Curve	Segmented Results
Segmentation

• Applied pretrained ProtTrans transformer model and trained a fully-connected classifier head.
• Achieved 98.438% accuracy on public leaderboard and 94.161% accuracy on private leaderboard. CNN baseline is 77.106% and 78.241%.

Trained U-Net models for semantic segmentation on the Cityscapes and Carvana datasets to build basic concepts on the U-Net architecture, with the following results:

• Carvana: 99.55% pixel accuracy, dice score of 0.9911 over the dataset downscaled to 320 x 480
• Cityscapes: 84.0% pixel accuracy, did not track mIoU

Performed object detection and tracking on videos using YOLOv8. Key features include:

• Static image and video object detection
• Instance tracking with third party library and custom algorithm respectively
• Car counting

Static Image Object Detection	Instance Tracking	Car Counting

A deep learning project exploring Generative Adversarial Networks (GANs) using the MNIST dataset:

• Experimented with different learning rates
• Observed the phoenomenon of mode collapse and the sensitivity of the GAN architecture to the learning rate
• Understanding the architecture of GAN, improvements made by WGAN, and also the principles of providing class conditions to GANs

Vanilla GAN	WGAN