Professional Experiences

LMM Humor Understanding, Research Assistant | Jun 24 - Present

Built a dataset of 30,000 comics through responsible web scraping (selenium and beautiful soup) whilst protecting identifiable information in the dataset.
Applied a combination of OCR and Computer Vision techniques like morphing (erosion and dilation) and image segmentation (adaptive thresholding) using OpenCV to separate the individual panels for annotations.
Designed and executed the experiments, methodologies and evaluations for both LMMs and humans for a comparative study.

NLP Research Intern | Aug 23 - Dec 23

Utilized quantized versions of state-of-the-art Large Language Models (LLMs) from HuggingFace such as Code Llama to analyze and detect code vulnerabilities in open-source software.
Explored and applied recent advances in research such as the “Graph of Thoughts” to encourage logical reasoning in LLMs, leading to significant improvements in detecting code vulnerabilities.
Extensively used tools and libraries such as LangChain to develop applications that integrate LLMs and perform Retrieval Augmented Generations (RAG).

Undergraduate Teaching Assistant | Sep 22 - Dec 22

Taught small groups (4-5 per group) of students in the 10.014 Computational Thinking for Design module.
Graded home assignments and lab work (programming a Raspberry Pi microcontroller) and maintained student records.
Assisted faculty with the preparation of 20 unique questions for upcoming classes that test students on their coding (Python) abilities - search algorithms, functions and object-oriented programming questions.

Projects

SmartDrive, Machine Learning Researcher (Computer Vision) | May 23 to Dec 23

Conduct literature reviews and identify relevant research and technologies for detecting driver fatigue.
Collect and preprocess sensor data from depth and thermal cameras to prepare it for algorithms to detect fatigue.
Implement and test algorithms relevant to Computer Vision, and utilized various techniques such as PERCLOS, Head Pose Estimation, and Eye Gaze Estimation.

Sentiment and Entity Analysis (Natural Language Processing) | May 23 to Aug 23
[Project Details]

Designed and applied a machine learning system using Hidden Markov Models to identify sentiments in informal sentences (tweets).
Tuned and validated the system by varying different smoothing methods which resulted in a ~240% improvement in the F scores (precision and accuracy) for identifying sentiments in the Spanish Language.
Instilled good practices for developing a machine learning framework to train and cross validate across different models to select the best model for optimal results.

Office of Student Life Web Application (Software Engineering) | May 23 to Aug 23
[Project Details]

Designed a web application for SUTD’s office of student life and used Amazon Web Services (AWS) such as Amazon Cognito, API Gateway and serverless Lambda Functions to deploy the web application.
Developed the application through a hybrid of Incremental and Iterative methodology and the Agile Manifesto whilst documenting the project with Unified Modeling Language (eg: Use Case diagrams, Sequential diagrams etc)
Performed unit, integration, and system tests for the backend functionalities of the application to achieve a coverage of 90.22% using Jest.
Integrated and configured AWS and Github for the team to develop and deploy the application (eg: AWS Identity and Access Management (IAM) for permissions, AWS Amplify for CI/CD deployment and AWS RDS for database management).

Solar Dehydrator DCH, Data Analyst | Sep 22 to Dec 22
[Project Details]

Worked in a group of five to design a unique and successful prototype to dehydrate food using Solar Panels, Arduino and a Fresnel lens which dried the (wet) test sponge within 1 hour. My main role is to analyze and interpret Solar Irradiance data.
Conducted data cleaning by fixing structural errors and filtering outliers (the 29th of February in the leap year).
Performed Q-Q plots to assess if the data fits a normal distribution. Subsequently, utilized the Central Limit Theorem to establish a confidence interval. Ultimately, we concluded that the solar dehydrator receives enough solar irradiance to dehydrate food continuously with a 90% confidence level.