Plainly

The problem

Contracts, research papers and government letters are written to be precise, not readable, and the tools that help mostly live in the browser and send your text to a server. Plainly works in every app, from PDFs to email, and keeps the AI on the machine: it runs on an ordinary 8 GB laptop with no graphics card.

What I built

• Selection capture from any app through Windows UI Automation, with a clipboard fallback that restores the person's clipboard exactly, even on error
• A card that opens beside the text without taking focus: web search tabs in a private web view, an offline WordNet dictionary, and Explain with AI
• Local explaining and translating through llama.cpp, with a GBNF grammar built per request so a small model can't drift into the wrong script
• A meaning check that compares numbers, dates and names between the text and the answer, and says on the card when one was dropped or invented
• An evaluation harness that runs 28 passages through each candidate model in CI; it picked Gemma 4 E2B (5.8 s median per rewrite on a 4-core CPU, 3.9 GB peak)
• A word book with history and CSV and Anki export, signed self-updates, and an installer built, smoke-tested and published on every push

Outcome

• A free preview for Windows 10 and 11, tested on a real PC and fixed from what that test found
• 201 Rust tests, 37 unit tests and 118 Playwright checks, with WCAG 2.2 AA scans in light and dark, before every push
• Nothing you read leaves the computer unless you open a web tab, and Offline only turns those off