logo
Third Year Computer Science Student

Hello! My name isBryan Melvison

Passionate about the application of new and emerging technologies, follow along the journey in learning with an open mind.

About Me

Motivated Lifelong Learner

Throughout my time in University, I've had the opportunity to equip myself with various fundamental knowledge revolving around Computer Science, such as Operating Systems, Database Management Systems, Algorithms, many more!

Bryan Melvison
+852 5136 7171
melvisonbryan@gmail.com
Major in Computer Science, with a Minor in Finance
Football, Badminton, Movies, Games, Travelling
Language Skill
English, Indonesian, Mandarin Chinese

My Interests

Data Science & ML

I love exploring and uncovering hidden patterns to gain insights from vast amount of data to come to a decision/solution with a strong basis.

Generative AI

I love the changes that GPT models have brought about to innovations. Currently doing research on the topic to gain a much better understanding of the topic.

Backend

I enjoy implementing the backbone of applications, to ensure smooth data flow, efficient processing, and secured interactions.

My Experiences

HKU Innovation Wing
Part Time Research Assistant (Robotics)
Oct 2023 - Dec 2023

  • Developed computer vision algorithms in OpenCV to perform color detection, contour extraction, and 2D coordinate mapping of colored objects to enable robotic arm calibration for pick and place operations.

  • Maintained and optimized existing code for an autonomous rover bot by refactoring and debugging to improve performance and precision.

Kodifly Limited
AI Engineer Intern
Jun 2023 - Aug 2023

  • Developed a cutting-edge Web Visualizer to facilitate real-time hosting of live camera feeds, live point clouds from the LiDAR Sensor, on-demand SLAM, and alarm system detection using React js, three js, and ros.

  • Built a robust backend infrastructure utilizing the ROS communication protocol, rosjs, and seamlessly integrated Livox SDK API for its detection capabilities on the Web Visualizer.

  • Calculated and retrieved the IMU of the LiDAR within the LiDAR Simulation with perfect accuracy rate, enabled the creation of a realistic SLAM map of the simulated landslide, and moving cars.

  • Successfully calibrated the live camera feeds result and overlay them on top of the point cloud feed based on theintrinsic and extrinsic parameters of the camera.

Highlighted Projects

Open Domain Question Answering System for RAG

Pytorch, Huggingface, MongoDB

  • Implemented a full Retrieval-Augmented Generation (RAG) pipeline for open-domain question answering, according to the Dense Passage Retrieval (DPR) paper, leveraging SOTA lightweight BERT models (ELECTRA, DistilBERT) for encoding passages
  • Fine-tuned model to achieve high performance metrics, with a recall of 0,95 and Mean Reciprocal Rank of 0,81
  • Managed data operations with MongoDB, ensuring robust data handling and quick retrieval for real-time question answering
View on GitHub

Decoder-only Transformer

Pytorch

  • Reconstructed the decoder only transformer model from the paper 'Attention is All you Need'
  • Trained on Shakespeare's work, it was able to autoregressively generate texts that mimic Shakespearean English, can also be trained on other corpus of texts
View on GitHub

PocketVocab by Purple Cow

React Native, Flask, Firebase

  • Made the quiz section of the app and the overall backend infrastructure using React Native for the UI and Python Flask for the server-side logic
  • Integrated Google Cloud Vision API for object recognition and annotation, and Google Cloud Translate API for translating object names to the user's selected target language, Gemini API for generating quizzes questions
  • Implemented Firebase services, including Firestore for database management, Functions for implementing serverless backend functionality and Storage for storing pictures taken by the user
View on GitHub

Intelligent Course Management System

OpenCV, Flask, MySQL, React.js

  • Developed facial recognition system to enable biometric authentication for course management platform
  • Built a machine learning pipeline including data collection, training, and integrated the model with the backend
  • Front-end and backend integration using Flask and MySQL for various functions
View on GitHub

Digit Recognition with CNN Model

Pytorch

  • Applied data augmentation techniques to prevent overfitting
  • Achieved a training accuracy of 94,3%, and testing accuracy of 94%
  • Modeled with 16 layers including convolution, pooling, activations, fully-connected, and dropout
  • Utilized max-pooling layer to effectively reduce dimension of feature map, numbers of parameters, and amount of computations performed to under 15 minutes
View on GitHub

Stock Predictor and Visualization with LSTM Model

Keras, Streamlit

  • Designed and implemented an intuitive user interface using the Streamlit framework, along with yfinance to retrieve up-to-date stock data, and visualize findings of essential aspects of stock market analysis
  • Employed LSTM modeling techniques to forecast future stock prices based on the historical stock price, on a 80-20 split, achieving an RMSE of 0,839
View on GitHub

Big Two

Java

  • Employed a multi-threaded approach to efficiently handle networking tasks, utilizing Java sockets for seamless network multiplayer functionality
  • Implemented Object Oriented Approach to effectively manage intricate game mechanisms and rules
View on GitHub

Virtual Paint

Open CV

  • Utilized image processing theories to find the difference between current and previous frame, and drawing a bounding box around moving objects, acting as a pen
View on GitHub

I welcome the opportunity to discuss future collaborations. If you're interested, please feel free to contact me.