$ ls -lt ./posts/
writing
AWS, Python, applied AI, other stuff. Mostly things I found vaguely interesting.
8
posts
Building the House of Knowledge: The Thinking Archivist
Part 8 of a series on building a real RAG system with real data. Letting the model search for itself, and finding out that 'search again' isn't the same thing as 'search better'.
Building the House of Knowledge: Lost in the Stacks
Part 7 of a series on building a real RAG system with real data. Retrieval scored fine on lookup and fell apart on temporal questions. This is why.
Building the House of Knowledge: Choosing Your Scholars
Part 6 of a series on building a real RAG system with real data. Nova Lite against Haiku 4.5, on a test set built to stop me picking the cheap one on reflex.
Examining the Scribe: Building a Test Set That Doesn't Lie
Part 5 of a series on building a real RAG system with real data - building an evaluation set honest enough to compare two models without fooling myself.
Teaching the Scribe: Automating D&D Session Notes with Claude
Part 4 of a series on building a real RAG system with real data - a Claude skill that enriches D&D session notes and improves the data underneath.
Opening the Doors: Serving the House of Knowledge over Cloudflare Tunnel
Part 3 of the House of Knowledge series. The RAG system works locally — now to make it accessible from anywhere. Without port forwarding, a VPS, or spending more money.
Building The House of Knowledge: The Bedrock Converse API
Part 2 of a series on building a real RAG system with real data - The Bedrock Converse API. A bit technical.
Building The House of Knowledge: RAG on a Budget
Part 1 of a series on building a real RAG system with real data - cheap as possible, no fancy GPU, hardly any RAM and to be honest totally unsuitable hardware.