AI/ML Engineering Portfolio

Welcome to my portfolio!

About Me

Hi, I am Ji Hin! I am an aspiring AI/ML engineer passionate about building intelligent systems and exploring how artificial intelligence can be leverages to optimise efficiency.

My work spans across machine learning, AI application development and software engineering. This portfolio showcases my projects, experiments in the AI and ML engineering field.

Projects

EasyAgents - Agentic System Visualiser

A visualiser for prototyping Agentic Systems for Google ADK. Wire up orchestrators, sub-agents, and evaluators, configure their prompts and tools, then submit a task and watch it execute in real time with live logs and a flamegraph-style timeline. Supports built-in templates or build-your-own from scratch.

Google ADK Gemini Multi-Agent System FastAPI

Pill Identifier

Classification and retrieval pipeline for loose pills. Uses computer vision models (YOLO, ResNet) and neural networks (RNN) and to identify pill imprint, shape, form, color to classify pills. Through a custom similarity function, top 3 most similar pills are retrieved. Model architecture based on Heo et al. (2023) in the Journal of Medical Internet Research.

YOLO ResNet RNN Machine Learning

Twitter Bot Detection

Machine learning pipeline to distinguish Bot users from genuine users based on a given dataset of Twitter accounts as part of a Kaggle Competition. This pipeline and methodology enabled me to place first among 117 other competitors in the Kaggle Competition. Key strategies include feature engineering by fine-tuning a Large Language Model and rapid iterative testing.

Fine-tuning LLMs Feature Engineering Fraud Analysis Machine Learning

Bitcoin Fraud Detection

Machine learning pipeline to predict fraudulent Bitcoin transactions by evaluating and comparing graph-based methods against traditional machine learning classification algorithms. Experiments and pipelines aimed to compare GNN performance against conventional classifiers like XGBoost, assessing whether relational information provides tangible improvements in fraud detection and operational effectiveness. Ran on Elliptical Dataset.

Graph Convolutional Network Pseudo-labelling Temporal Analysis Fraud Analysis Machine Learning

Local RAG

Locally hosted RAG assistant for MATLAB-syntax Q&A on a single RTX 3060 Ti 8 GB setup.
Created my first RAG that uses MATLAB documentation as the knowledge base. Models are downloaded from HuggingFace and Quantized using BitsandBytes. Some basic data preprocessing with PyMuPDF was implemented on a separate pipeline. ChromaDB was selected as the vectorstore.

Python LangChain HuggingFace ChromaDB