top of page

Portfolio

This is my Portfolio Page.

Here are my SaaS projects:

  • DenseReader.com — AI auto-summarizer using chain-of-density for multi-level summarization.

  • Osmomark.com — AI auto-highlighter and study partner that can hide relations/meanings for active review.

  • SaaSique.com — AI market-validation tool that scans 10 channels for pain points, improvements, competitors, and pricing before you build.

  • Marudio.com — AI audio volume enhancer and denoiser.

  • H2ONote.com — AI water-drinking tracker and reminder for daily hydration health.

  • LobsterKet.com — AI marketplace where owners' agents autonomously trade items and services between each other.

  • Quomark.com — AI-facilitated team decision-maker for chat, voting, and tiebreaking.

  • TuneSoar.com — AI that reads your current task and serves matching binaural beats for focus, relaxation, or memorization.

  • VeryWhy.com — Conversational AI using rehab and psychology to surface deep motivations for positive life change (quitting smoking, exercising, studying).

The following are my other demos:
 

🎬 MovieMatch - AI Movie Recommendations
Neural Collaborative Filtering System | 66.2% Improvement Over Baseline

Try my PyTorch-powered recommendation engine trained on MovieLens 100K. Enter a user ID to get personalized movie suggestions.

📈 StockForecast - AI Time Series Prediction Dual-Architecture ML System | <10% MAPE with 20-Second Training

Try my TensorFlow-powered forecasting dashboard featuring LSTM + Prophet models. Enter a stock ticker to get real-time price predictions with confidence intervals.

🏥 HK Healthcare RAG Chatbot - Production AI System | 66.2% Cost Reduction vs Traditional Cloud

Try my Cohere-powered RAG chatbot trained on 616 Hong Kong healthcare documents. Ask healthcare questions to get AI-generated answers with source citations.

🤖 Multi-Agent AI Research Assistant - Production AI System | $0/Month Serverless Deployment

Try my CrewAI-powered multi-agent system with 3 specialized AI agents (Researcher, Writer, Critic). Enter a research topic to get an automated, comprehensive research report with live web search and citations.

🎨 PaliGemma Image Captioning | Vision-Language Model Fine-tuning (Down for maintenance)

Upload fashion images to get AI-generated captions describing clothing items, colors, seasons, and styles. Fine-tuned PaliGemma-3B model specialized for fashion e-commerce product descriptions. Supports PNG, JPG, GIF, WEBP.

Update — July 2026: Retrained via a full port to the ms-swift framework, running on a single free Kaggle T4 GPU. The new LoRA fine-tune reached an eval_loss of 0.279 — outperforming the original transformers baseline (0.472) while cutting training time from ~9 hours to 1h 34m. Peak GPU memory stayed at 14.4 GB, comfortably within a 16 GB T4. Captions cover article type, colour, season, and usage with ~90% token accuracy on held-out validation data.

Link: https://github.com/Donald8585/paligemma-fashion-lora-msswift

Credit Card Fraud Detection | Real-Time ML System

Interactive demo for credit card fraud detection with XGBoost. Analyzes transactions in real-time with 99.9% accuracy across 8 realistic scenarios.

Diabetes Risk Prediction | Real-Time ML System

Interactive demo for diabetes risk prediction with XGBoost. Analyzes patient health data in real-time with 68% accuracy across 26 clinical features.

Warehouse Robot Navigation | DQN + HER + Curriculum Learning

Interactive demo comparing BFS pathfinding vs. Deep Reinforcement Learning agent. Trained using DQN with Hindsight Experience Replay across 6 difficulty levels with 300k training steps, achieving 100% success on easy tasks.

Stablecoin Agent | DeepSeek V4 + viem RPC + React/Node.js

Interactive on-chain assistant that reads live USDC/USDT balances and transfers on Ethereum via blockchain RPC. Built with a React frontend, Node.js/Express backend, and a DeepSeek V4 Flash agent that routes natural-language questions to viem RPC tools (balanceOf, getLogs), exposed as an MCP server and validated with an LLM-as-judge eval harness.

bottom of page