SwarnaSabd
A desktop text refiner for local and hosted language models.
Why it exists
01I write notes in a mix of English and Hindi, then rewrite them three times before they sound right. I wanted a tool that does the boring passes: tighten, clarify, fix grammar, keep the meaning.
I did not want to paste personal writing into a random website. SwarnaSabd can run entirely against a local model, with a hosted model as the faster option.
Recorded 2026
Kaushambi, India
The constraints
02- 01Windows was the only target; packaging for macOS would have needed a Mac.
- 02The app had to work without an internet connection.
- 03A 75MB portable build was the limit I set for distribution.
The approach
03A pipeline, not a single prompt
The text goes through separate passes: preserve meaning, tighten sentences, fix grammar, then compare against the original. Each pass can use a different model.
Local-first, hosted when asked
Ollama is the default. A hosted model is opt-in and the UI states when text leaves the machine.
What it actually does
04- 01Multi-pass refinement with a visible diff.
- 02Local Ollama models and hosted model options.
- 03Portable Windows build with auto-update through GitHub releases.
- 04Presets for notes, emails and longer documents.
Technical notes
05Electron shell, React renderer, and a small Python bridge for prompt orchestration. Releases are built with electron-builder and published through GitHub Actions.
The first version shipped without a diff view. That was a mistake: users could not see what the model changed. Version 1.1 added a side-by-side comparison.
- Role
- Design, Electron app, prompt pipeline, release packaging
- Year
- 2026
- Platform
- Windows
- Stack
- Electron, React, Python, Ollama, electron-builder, GitHub Actions
- Status
- Shipped
- Links
- Download ↗
Honest limits
06- 01Windows only.
- 02Quality depends heavily on the selected model; small local models flatten nuance.
- 03The portable EXE is not code-signed, so SmartScreen warns on first launch.
- 04No support for PDFs or scanned documents.
What it taught me
07“Shipping a desktop app is mostly distribution: code signing, updates, antivirus warnings and file paths. The model was the easy part.”