AI, in Depth
Here I write about how large language models (LLMs) actually work, for anyone who uses AI tools and wants to understand what is happening underneath. That means the building blocks: tokenization, context windows, prompting, and the difference between a model and the harness around it.
A model like Claude, GPT, Gemini or Gemma is only part of the picture. Harnesses such as Claude Code, Cursor and OpenCode decide what the model sees, which tools it can call and how its work is checked. Platforms like OpenRouter give one way to reach many models, and tools like LM Studio let you run a local LLM on your own machine. Choosing between local and API models is a real trade-off in cost, privacy and quality, and I cover it here too.
My own work in this area includes custom Claude skills, a multi-agent Claude Code pipeline that audits a full production codebase in under 6 minutes, and an AI chatbot for Messenger and Instagram that runs on OpenRouter. I also co-developed Tell Liora, an offline mobile app that pairs an on-device Gemma model with 8 WHO decision tables.
Posts here aim to leave you understanding something better than before: plain explanations first, with the technical details and sources for anyone who wants to go deeper.
Start here
- What Is Tokenization, and Why Does It Decide What AI Costs?
Start with tokens, the unit every model reads, writes, limits and charges by.
All posts on AI, in Depth
What Is Tokenization, and Why Does It Decide What AI Costs?
Every AI model reads, writes, limits and bills you in tokens, not words. Once you understand tokens, AI pricing and limits start to make sense.
Related projects
- Tell Liora: An offline pregnancy care app that runs Gemma on the phone, built at the AppBuildersPH Hackathon 2026.
Get new posts by email
One email every two weeks with new posts, news takes and one personal note. No spam, unsubscribe any time.