TwinChat

A pipeline that ingests my blog posts, essays, and drafts; generates a fine-tuning dataset; trains a small model on Modal; and serves the resulting chat widget on this site, streamed through a Netlify function. The point isn't accuracy — it's voice.

Talk to it:

Talk to my twingemma 4 fine-tune (e2b · LoRA r=128) · with retrieval

How did you build this?What does the training set look like?Why a fine-tune instead of a system prompt?When does it sound the least like Simon?

↵