I ship products end-to-end — from infrastructure to interface. I built AI Agent Folio, a multi-app SaaS platform running entirely on Oracle's Always Free tier, hosting 12+ apps on a single ARM VM with local LLM inference, full e-commerce, s…
A complete multi-app SaaS platform running on a single Oracle Always Free ARM VM (4 OCPU, 24GB RAM). 12+ subdomains, unified auth gateway, cross-app messenger, community forum, and local LLM inference via Ollama. Zero monthly hosting cost.
OpenSingapore HDB resale price predictor trained on 12,000+ real transactions from data.gov.sg. MLP regressor with 95% confidence intervals, integrated with OneMap reverse geocoding and OpenStreetMap amenities. R²=0.789, MAE ~$50k.
OpenSingle sign-on for all platform apps using JWT cookies signed at the apex domain. PBKDF2 password hashing, stateless CSRF tied to session, password reset via one-time tokens, 14-day free trials, admin panel.
OpenSingapore restaurant and attraction discovery, powered by Google Places API (New). Filterable by category, cuisine, dietary, price level. Map-based browsing with cost-controlled crawl staying under Google's free tier.
OpenFloating chatbot on every authenticated page, powered by Ollama (qwen2.5:7b) running locally. Agent-style responses with markdown links that drive users to the right app. No third-party AI API calls.
OpenModern storefront with Stripe checkout, free-shipping threshold UX, image gallery, cart persistence via signed cookies, and an admin panel for product CRUD.
OpenCode generation, review, and explanation tool with image-to-code (vision) support. Streams responses from qwen2.5-coder:7b and qwen2.5vl:7b via Ollama, with fallback to deepseek-r1:7b.
OpenAtomic SQLite backups of all 8 application DBs to Oracle Object Storage every night, with 2-day retention and gzip compression. Restore tooling included.
Click any spot in Singapore — get historical resale price ranges plus an AI-predicted value with 95% CI. Trained nightly on the freshest HDB transactions from data.gov.sg.
Try itLocal LLM tutor — explains concepts at any depth from elementary to PhD level. Powered by qwen2.5:7b via Ollama, with conversation history in browser local storage.
Try itEquity research assistant combining market data and LLM analysis. Pulls fundamentals, generates research reports, simulates options strategies (educational use only).
Try itSecurity operations chatbot for triage, log analysis, and incident summarization. Local-only inference for data privacy.
Try it
OpenAI Whisper transcribing live audio to text in real time. Demonstrates automatic speech recognition (ASR) with multilingual support and word-level timestamps.
🔗 Open link
Retrieval-Augmented Generation pipeline: documents chunked, embedded with nomic-embed-text, retrieved by cosine similarity, then answered by a local LLM. No hallucination on domain facts.
🔗 Open link
OpenAI CLIP matching images to text prompts with no task-specific training. Shows contrastive vision-language alignment: input any image, output natural language label ranked by confidence.
🔗 Open link
StyleGAN2 generating photorealistic synthetic faces. Visualises the generator/discriminator adversarial training loop and latent space interpolation between identities.
🔗 Open link
Deep Q-Network agent learning to play Breakout from raw pixels with no hand-coded rules. Shows reward curve convergence over 1M training steps on a standard gym environment.
🔗 Open link
YOLOv8 detecting and tracking multiple object classes at 30+ FPS on standard video. Demonstrates bounding box regression, class confidence scores, and multi-scale feature extraction.
🔗 Open link
SSL label propagation on MNIST with only 100 labelled samples out of 60,000. Graph-based pseudo-labelling achieves ~94% accuracy, visualised with t-SNE cluster evolution across epochs.
🔗 Open linkAnnotated code, experiments, and analysis — click any card to read inline.
End-to-end ML pipeline: advanced feature engineering, EDA with correlation heatmaps, SHAP analysis, and HDB resale price…
Comparing Isolation Forest vs LSTM Autoencoder on MNIST anomaly detection. Includes precision-recall curves, threshold t…
PPO agent solving CartPole-v1 in under 200 episodes, then MLP policy network on LunarLanding-v2. Covers actor-critic arc…
Fine-tuning BERT-base on sentiment classification. Covers tokenisation, attention masks, learning rate warmup, confusion…
Explainable AI on Singapore COE bidding data: SHAP values, temporal trend analysis, price elasticity modelling, and comp…
Transfer learning from ResNet-50 on custom dataset, CLIP zero-shot classification, and GAN image generation. Includes Gr…