attention.sh: why we exist

attention.sh is the applied AI research division of Jinacode Systems. This site is where we publish what we learn building large language model systems — blogs, paper notes, and research write-ups.

What we work on

Three threads run through everything we do:

1. Transformers, understood from first principles. Not just calling APIs — the math of attention, distillation, and scaling laws, because the trade-offs in production are invisible without it.
2. Fine-tuning that earns its keep. Parameter-efficient methods (LoRA, distillation) applied where a specialized model genuinely beats a general one — with evaluation infrastructure to prove it either way.
3. Sovereign LLMs. Helping organizations build models they own: their data stays inside their boundary, the weights are theirs, and the system runs on infrastructure they control. For regulated domains — tax, law, finance — we think this is not a preference but a requirement.

Where the research lives

Right here, on this blog. We publish what we learn as we go — paper notes, evaluation findings, and write-ups from real deployments (an Indian tax-law chatbot among them). The deeper study material behind it stays in our internal lab notes; the parts worth sharing get distilled into posts.

The research page tracks what’s in progress. First publications are compiling.

← all posts