Projects & Case Studies

Interactive Web App: A dynamic Science & Math quiz for Primary 5 students. Pulls 5 fresh questions on each play from a public trivia API and scores you in real time.

• Dynamic Content: Fetches a new random set of questions each round from the Open Trivia DB API — no two runs are the same.
• Client-Side Logic: Fisher–Yates answer shuffling, instant feedback, live scoring, and a retry flow — all in vanilla JavaScript.
• Zero Backend: Runs entirely in the browser, served as static files from the same CloudFront distribution as this site.

This Website: A modern, responsive personal portfolio built entirely on AWS serverless infrastructure.

• Infrastructure: Automated provisioning via Terraform (S3 & CloudFront) with custom modules.
• Secure Authentication: Eliminated long-lived Access Keys (AK/SK) by implementing GitHub OIDC to assume AWS roles with short-lived credentials for Terraform runs.

Event-Driven AI: An automated pipeline using Amazon Bedrock (Claude 3 Haiku) to summarize and extract insights from PDF documents uploaded to S3.

• AI Orchestration: Lambda-based workflow using urllib.parse to handle URL-encoded S3 events.
• Data Persistence: Extracts structured JSON from unstructured text and stores results in DynamoDB.

Private RAG Tool: A 100% offline tool that allows secure chat with sensitive PDF/Excel documents using local LLMs, ensuring zero data leakage to the cloud.

• Tech Stack: Built with Python, Streamlit, LangChain, and Ollama.
• Benchmarks: Tested against Llama 3 (8B) and Mistral-Nemo (12B). Mistral-Nemo proved more accurate in extracting entities from Spanish legal docs.
• Privacy: Uses ephemeral in-memory vector storage (ChromaDB) to ensure data is wiped instantly after analysis.

Learning by Building: A five-step lab that builds up from a single Claude API call to a working AI agent, adding exactly one concept per step — written to make RAG and MCP concrete rather than abstract.

• The Agent Loop: Written by hand first — request, execute tool, feed result back, repeat — before handing it to the SDK's tool runner. Every agent framework is a wrapper around this.
• RAG From Scratch: Load, chunk, embed, cosine search — no vector database and no LangChain, so the mechanics stay visible. Embeddings run locally via Ollama, with a pure-Python TF-IDF fallback.
• Model Context Protocol: A standalone MCP server exposing the retrieval tools, discovered over JSON-RPC by the agent — and by Claude Code, unchanged.

Roadmap: Enhancing the static portfolio with dynamic capabilities using a Serverless architecture.

• Implementing API Gateway as the REST entry point.
• Developing AWS Lambda functions (Python) to process contact forms and dynamic content.
• Integration with DynamoDB for data persistence.

Infrastructure as Code: Built a reusable Terraform module to automate the deployment of secure, high-performance static websites.

• Provisions ACM Certificates for SSL/TLS.
• Configures CloudFront Distribution with default cache behaviors.
• Secures S3 Buckets with OAI/OAC policies.