Longhand: The short way to file your tax abroad

Longhand is an auditable AI tax assistant built for US citizens working PAYE jobs in the UK — a group stuck between HMRC and the IRS with no tool that clearly shows its work. The core problem: US expats must choose between the Foreign Tax Credit (Form 1116) and the Foreign Earned Income Exclusion (Form 2555), and picking wrong costs real money. FTC works better when UK taxes are high; FEIE works better when they're low. Existing tools either bury this decision in generic advice or use an LLM to "do the math," which is a compliance risk when the numbers touch the IRS. Longhand's core architectural principle is that the LLM never calculates numbers. Users enter UK salary, tax paid, filing status, days abroad, and dependents. A deterministic Python tax engine computes the FTC vs FEIE comparison, recommends a route, estimates US tax impact, and flags required filings (Form 1116, Form 2555, FBAR/Form 8938). Only after the math is done does the LLM step in — generating a plain-English explanation grounded in the calculation trace and cited IRS/HMRC/treaty snippets pulled from a curated RAG store. Every number in Longhand can be traced back to a deterministic function, not a model guess. That auditability is the product's core differentiator, not a bolt-on feature. Built with a FastAPI backend, a Next.js/React frontend, and inference on AMD MI300X (via vLLM) with a Fireworks AI/Gemma 3 27B Instruct fallback, all running in Docker Compose. The demo covers the PAYE-only salary earner case, with a full calculation trace viewer and filings panel so users can see exactly why the tool recommends what it does.