← All work
live Author & Engineer 2026

Simasia

A brand-tone guardrail that scores whether an AI reply sounds like you — and steers it until it does.

Python scikit-learn Embeddings CLI PyPI

LLMs are fluent but tone-blind. Simasia adds the missing judge: freeze a large embedding model, train one small logistic-regression head on a brand’s voice, and you can rate any reply from 0 to 1 for a few kilobytes of weights per brand.

Because the judge is a classifier, not a writer, everything else follows cleanly. It explains a score with the closest on-brand and off-brand examples — no language model needed. It can train from raw text, a file, or a list of URLs, and even manufacture off-brand negatives from on-brand copy alone. And it closes the loop: hand it a generator and it rewrites until the reply clears your threshold. Shipped as an installable package with a config-file CLI for people who don’t write code.

Technical Deep Dives

  • λ
    Simasia: A Brand-Tone Guardrail for LLM Output

    How to build a small, cheap classifier that scores whether an AI reply sounds like your brand — and then steers the model until it does. One frozen embedding model, one logistic head per brand.

    Jul 2026 / Advanced
← Back to all work