A local AI that never sends a client's return anywhere

The problem

Every return went through an hour of manual redaction before it touched Claude, because there was no way to be sure client Social Security numbers and identifiers were actually safe, and Section 7216 makes that exposure criminal, not just embarrassing. On top of that, long chats drifted: a Rule 83B filing timeline that had been carefully built came back with its dates scrambled a week later, for work where dates are the whole job.

At roughly $2,000 in AI credits every three weeks, none of it routed to a cheaper model for the routine parts, the bill was also just wrong for the work being asked of it.

What we proposed

Three architectures were laid out with honest tradeoffs: a fully local machine that never sends a byte of client data to any outside AI provider, a hybrid split where only client-free legal research goes to the cloud, and a redaction gateway that strips identifiers before anything reaches Claude. The recommendation was the fully local option, because the exposure here meant the bar was airtight or nothing.

Nothing gets bought on faith. The plan opens with a hosted bake-off, the exact recommended model tested side by side against Claude on synthetic returns, for pennies, so the quality gap gets judged firsthand before any hardware purchase. Only if that passes does a proof of concept run on the actual machine, inside the return window, so the purchase stays reversible even then.

The architecture is grounded against the specific failure that started this: primary tax authority gets cited from a maintained local library instead of memory, all arithmetic runs through a calculator tool instead of the model's head, and every fact and date gets extracted once into a structured timeline the AI must defer to on every later analysis, the direct fix for the scrambled 83B dates.