Desktop Autocorrect — Case Study

Challenge

Desktop operating systems lack built-in, phone-style autocorrect that works everywhere you type — Notepad, browsers, IDEs, and chat apps. The goal is a privacy-first, offline utility that catches typos at word boundaries without sending keystrokes to the cloud or correcting passwords and secure fields.

Solution Architecture

Global input capture — low-level keyboard hooks with platform abstraction (Windows, macOS, Linux)
Correction pipeline — buffer → policy check → spell engine → inject or suggest → learn
Spell engine — dictionary lookup, Levenshtein edit distance, word-frequency ranking, built-in typo rules
Smart policy — skips URLs, emails, file paths, camelCase/snake_case, and password/secure fields
SQLite learning — user dictionary, self-correction memory, reject/undo tracking
Windows optimization — native Win32 Edit control replacement for Notepad/WordPad (avoids fragile backspace injection)
MVP UI — CustomTkinter dashboard, system tray controls, hotkeys (toggle, undo, manual suggest, emergency stop)
Future path (in progress) — Rust core engine (autocorrect-core), Tauri 2 shell, native OS text services (TSF/IBus)

flowchart LR
  Keys[Keystrokes] --> Hook[Platform Hook]
  Hook --> Buffer[Word Buffer]
  Buffer --> Policy[Policy Engine]
  Policy --> Spell[Spell Check]
  Spell --> Act[Inject or Suggest]
  Act --> Learn[(SQLite Learning)]
  Tray[Tray / Dashboard] --> Config[Config and Modes]

Results

Shipped working Python MVP: global hooks, auto/suggest/off modes, per-app profiles, and distributable Windows .exe via PyInstaller
Built 88+ automated tests covering spell engine, policy, learning, and Windows edit injection
Documented architecture, security threat model, and 7-phase roadmap in dedicated docs/ folder
Phases 5–7 scaffolded: Rust core crate, Tauri 2 shell, native OS integration stubs, optional local AI ranker

Timeline

| Phase | Focus | Status | |-------|-------|--------| | 1 | Global hook, word buffer, dictionary autocorrect, tray, config | Complete | | 2 | Suggestion overlay, suggest/auto/off modes, undo hotkey | Complete | | 3 | Secure-field skip, URL/email/path exclusions, ring-buffer logging | Complete | | 4 | SQLite learning, locale dictionaries (en, es) | Complete | | 5 | Rust core + Tauri UI | In progress | | 6 | Native OS text services (TSF, IBus) | Planned | | 7 | Optional local AI ranking | Planned |

Challenge

Solution Architecture

Global input capture — low-level keyboard hooks with platform abstraction (Windows, macOS, Linux)

Correction pipeline — buffer → policy check → spell engine → inject or suggest → learn

Spell engine — dictionary lookup, Levenshtein edit distance, word-frequency ranking, built-in typo rules

Smart policy — skips URLs, emails, file paths, camelCase/snake_case, and password/secure fields

SQLite learning — user dictionary, self-correction memory, reject/undo tracking

Windows optimization — native Win32 Edit control replacement for Notepad/WordPad (avoids fragile backspace injection)

MVP UI — CustomTkinter dashboard, system tray controls, hotkeys (toggle, undo, manual suggest, emergency stop)

Future path (in progress) — Rust core engine (autocorrect-core), Tauri 2 shell, native OS text services (TSF/IBus)

flowchart LR Keys[Keystrokes] --> Hook[Platform Hook] Hook --> Buffer[Word Buffer] Buffer --> Policy[Policy Engine] Policy --> Spell[Spell Check] Spell --> Act[Inject or Suggest] Act --> Learn[(SQLite Learning)] Tray[Tray / Dashboard] --> Config[Config and Modes]

Results

Shipped working Python MVP: global hooks, auto/suggest/off modes, per-app profiles, and distributable Windows .exe via PyInstaller

Built 88+ automated tests covering spell engine, policy, learning, and Windows edit injection

Documented architecture, security threat model, and 7-phase roadmap in dedicated docs/ folder

Phases 5–7 scaffolded: Rust core crate, Tauri 2 shell, native OS integration stubs, optional local AI ranker

Timeline