How Many R's in Strawberry? Training Echo, My Own LLM
I trained Echo, a 1.38B-parameter GPT, from scratch on 8 H100s using nanochat and Claude Code. She beats 2019's GPT-2 on CORE for $69, after $104 of failures.
A collection of more project based articles where I produced real outputs.
I trained Echo, a 1.38B-parameter GPT, from scratch on 8 H100s using nanochat and Claude Code. She beats 2019's GPT-2 on CORE for $69, after $104 of failures.
How I built an AI Product Decision Agent that researches product questions and produces structured decision briefs.
I built a second AI agent for my site with Claude Fable 5 in one ~10-minute session: a chatbot that answers from my Goodreads reviews, before the model was pulled.
You can just do things - going from a free WordPress blog to a self-managed site using only natural language
Bolt and Lovable go head-to-head building SpacEd, a space-facts app for young children, with my five-year-old as judge.
ChatGPT, Claude, Gemini, and Lovable get the same vague prompt to build a portfolio site, and the winner goes live on Vercel.
Calling foundation models through the Bedrock API with fewer than 100 lines of Python, then wrapping them in a Streamlit front end.