My Projects (ML / Software)

Awarded first place globally as part of Alibaba's flagship AI hackathon, putting Singapore’s name on the global AI innovation stage.

Comprises: LLMs for product attribute extraction, product classification, and RAG for business queries on export compliance.

Featured by many Chinese news agencies, Yahoo Finance, and SMU News.

Changes the background and style of any single-person or group photos uploaded.

Comprises: DETR-ResNet50 for panoptic segmentation + background change, and PyTorch DeepDream and style transform.

Showcased at multiple university events (Metamorphosis 2024, Vivace 2024, SMUBIA Sessions 2024).

My first ever machine learning project: matching input user’s resume to the most suitable current online job listings.

Comprises: web-scraper for job information data, and an ensemble of TF-IDF and transformer-based embedding models.

Outperforms traditional keyword-based matching systems.

Web application handling players, hosts, clubs, administrators, and tournaments for football in Singapore.

Comprises: Java Spring Boot backend with 4 microservices, including clubs, users, tournaments, and a chatbot microservice.

Developed with Agile (Scrum and Xtreme Programming) practices in React.js frontend.

System that detects questions in search bar and returns LLM-generated answers for quick query resolution, then suggesting more highly-relevant videos to the query.

Comprises: Support Vector Machine (SVM) to classify questions, LLMs to generate answers, mPLUG-2 for video captioning and semantic matching.

Search query relevance increased from ~25% to ~80%.

Developed a web application to help keep track of upcoming dividend payout for dividend capture trading strategy.

Filter by non-bearish stocks or positive EV based on backtested data over the past 20 years on performance of the strategy.

Suggestion for buy / sell dates and target and stop losses based on backtested information.

Application that allows a user to play the popular card game, Hearts.

Complex logic implemented for computer players, alongside smooth frontend animation and card rendering.

Built using Java backend.

Developing a coding assistant that can generate and debug code more effectively than the base LLM.

Using LoRA to fine-tune the base LLM more cost-effectively by injecting low-rank matrices into the transformer architecture and only updating those weights.

Photos and GH to be updated soon: Stay tuned!

** Still ongoing ***

Finalist in AWS x VCT Hackathon.

LLM-based assistant to retrieve competitive Valorant player information and form teams according to their playstyle, region, and tier.

Leveraged Amazon Bedrock agents and knowledge bases to build low-code RAG agents and vector stores. (utilising Cohere Embed English 3, Claude 3 Haiku)

Photos and GH to be updated soon: Stay tuned!

** Still ongoing ***