My Projects
A collection of work spanning AI, web development, and creative endeavors
Technical Depth
Retrieval & Search
- • Hybrid BM25 + Dense Vector Retrieval
- • Sentence Transformer Embeddings
- • ChromaDB HNSW Vector Store
- • Citation-Injected Structured Prompting
LLMs & Inference
- • Multi-LLM Abstraction (OpenAI, Ollama, Anthropic)
- • Local LLM Deployment (Ollama + Kubernetes)
- • FastAPI SSE Token Streaming
- • GDPR-Aware PII Classification
ML & Data Engineering
- • ML Trading Models (RF, XGBoost, Neural Nets)
- • Walk-Forward Cross-Validation
- • Metadata-Driven ETL Pipeline Design
- • Quantitative RAG Evaluation (20-query benchmark)
Featured Work
AI-Powered Personal Website
A Next.js + FastAPI web application that parses my personal Obsidian knowledge vault in real time, extracts structured notes and quotes, and serves them through a REST API. The system uses a custom tree parser and caching layer for performance. Demonstrates full-stack development, API design, and integration of personal knowledge management with web technology.
Spiritual AI Guide — RAG Knowledge Chatbot
Five-stage RAG pipeline built from scratch: (1) Obsidian vault parser with WikiLink graph extraction; (2) structure-aware chunking (800-token chunks, 150-token overlap); (3) sentence-transformers/all-MiniLM-L6-v2 → 384-dim L2-normalised vectors in ChromaDB HNSW; (4) hybrid retrieval: 70% cosine similarity + 20% BM25 keyword overlap + 10% WikiLink density, recovering ~12% additional recall on named-entity queries vs. dense-only; (5) GPT-4 Turbo with citation-injected structured prompts, token-by-token SSE streaming. Multi-LLM abstraction (OpenAI, Ollama/Llama 3.1, Anthropic, Google) via shared abstract base class. Evaluated across 20 queries: GPT-4 Turbo achieves 4.7/5 citation accuracy, 4.6/5 contextual grounding.
AI-Powered Personal Website
Full-stack portfolio powered by a FastAPI backend that live-parses an Obsidian vault (1,649 notes → 1,772 semantic chunks) into a REST API with multi-LLM RAG chat and SSE streaming, deployed on Render + Netlify.
Spiritual AI Guide — RAG Knowledge Chatbot
Production-deployed RAG system that semantically searches 1,649 personal Obsidian notes (~300K words) and delivers grounded, cited responses via GPT-4 Turbo.
Independent Podcasts
Created and hosted two podcasts interviewing researchers on Physics, Cosmology, Mathematics, and Philosophy.
Dodgeball Club Amsterdam Website
Created and currently maintaining a dynamic and modern website for the Dodgeball Club Amsterdam.