Client material that never leaves the building

Dor Iluz's property-law firm in Israel — the same client as Law Clock, Marmadbir and the firm's website. Hebrew throughout, installed and run by people who are not technical and should not have to be.

The firm wanted drafting help. It could not put client material into a cloud model to get it — not as a preference, but as a condition of the work. Any tool that solved the drafting problem by sending case files to an API was not a candidate.

• A drafting assistant that runs offline. Sources come from the firm's own library, the draft is written locally, and every authority is checked against the indexed corpus before the lawyer sees it.
• An installer for a non-technical user. Double-click, wait. It detects the hardware and picks a model to match, finds or installs Python without administrator rights, builds an isolated environment inside the project folder, installs or locates Ollama, downloads the model, indexes the library, creates a desktop shortcut and runs a self-test. Re-runnable at any time — every step skips what is already done.
• Templates for demand letters, pinui-binui claims and third-party notices, with DOCX export straight into the firm's workflow.
• A diagnostic script for when something breaks, so the first move is not a phone call.

Python (standard library BM25 retrieval) · Streamlit · Ollama · python-docx · pypdf · PowerShell installer. The Anthropic backend exists but is commented out of the requirements and off by default.

The offline promise shapes the architecture, not just a setting. Retrieval is BM25 written against the standard library rather than embeddings, because a dense-retrieval model means a model download — and the moment the tool needs to fetch something it is no longer straightforwardly offline. Formulaic legal Hebrew rewards lexical matching, so the constraint cost less accuracy than it looks like it should. Cloud drafting is present as an option the partners can turn on, and the config says in as many words that sending client material off-premises is their decision, not the software's.

Hebrew on Windows fights back. The launcher files are Hebrew-named one-line wrappers around English-named scripts. The Windows shortcut API goes through the system ANSI codepage, so any Hebrew character in a target path becomes a question mark on a machine that is not in a Hebrew locale. The split is deliberate: the shortcut always points at the English file.

No outcome measured yet; this study describes what was built and verified.