a division of jinacode systems

Transformers, fine-tuning, and sovereign LLMs.

attention.sh is the applied AI research division of Jinacode Systems. We work close to the metal of large language models — evaluation, retrieval, and fine-tuning — and help organizations build sovereign LLMs: models they own, understand, and run on their own terms.

▶ run the forward pass ⚙ the illustrated decoder

a 3D journey through a transformer — and how a decoder writes, one clear figure at a time.

./attention.sh --focus

→ fine-tuning & parameter-efficient adaptation (LoRA, distillation)

→ evaluation infrastructure for domain-specialized LLMs

→ retrieval & routing for production systems

→ sovereign LLM deployments — your data, your weights, your infra

12study notes
10source papers
8attention heads
∞tokens of curiosity

What we do

fine-tuning

Domain-specialized models via LoRA, distillation, and careful data curation — grounded in the math, not just the tooling.
evaluation

Eval harnesses that tell you what your model actually gets wrong, before your users do.
sovereign LLMs

End-to-end help building models you control — on-prem or private cloud, with no data leaving your boundary.
research

Paper deep-dives and applied write-ups from real deployments, published on the blog.

Latest writing

2026-07-16 attention.sh: why we exist

All posts → · Research →