Voice Dictation
Real-time speech-to-text with manual and auto-stop modes.
Healthcare
RxScribe — Voice-to-Prescription Desktop App
The doctor talks. The prescription appears. A desktop app with a local AI brain that turns 90 seconds of speech into a printable Bengali prescription — without ever touching the cloud.
100%
On-device inference
2
Languages (EN/BN)
0
Cloud dependencies
Screenshots
Overview
A privacy-absolute clinical tool: doctors speak, an offline AI listens, and a structured, printable prescription appears — in English or Bengali — in seconds. Whisper.cpp transcribes, a bundled LFM2.5 1.2B model parses, and not a single byte of patient data leaves the machine. Ever.
The Challenge
Doctors waste 30% of every consultation writing. Cloud EMRs are too slow, too expensive, and — for any patient whose data is genuinely sensitive — a legal landmine. The 'AI prescription' tools that exist all send patient transcripts to OpenAI. That's not just bad practice; in many jurisdictions it's illegal. The market needs a tool that's smart, fast, and provably private.
The Solution
I built RxScribe as a Wails desktop app with a complete on-device AI stack: whisper.cpp for speech, llama.cpp running a 1.2B-parameter model for transcript-to-prescription parsing, GBNF grammars to force structured output, and a gofpdf renderer with embedded Bengali typography. Validated with golden tests. Patient data stays in local SQLite. Works in a clinic with no internet, in a village with no electricity (battery permitting).
Tech Stack
Backend
GoWails v2SQLitellama.cpp supervisorGBNF grammars
Frontend
SvelteViteTypeScriptCustom design system
AI
Local LLM (llama.cpp)Tool registry patternGrammar-constrained generation
Speech
Cross-platform speech capture (Linux + Windows)
gofpdf with embedded NotoSansBengali
Key Features
AI-Assisted Parsing
Local LLM converts free-form transcripts into structured prescription fields.
Doctor & Patient Profiles
Persistent doctor settings, patient history, and prescription archive.
Bengali PDF Output
Print-ready prescriptions with embedded Bengali typography.
Fully Offline
All inference, storage, and PDF generation happen on-device.
History & Search
Browse past prescriptions and patients with full-text search.
My Role
Founding Engineer
Sole engineer — owned product design, Go backend (store, speech, PDF, LLM supervisor, tool registry), Svelte UI, and the local-LLM integration with golden-test validation.
Results & Impact
- Doctors get printable prescriptions in seconds without leaving their desk
- Local LLM keeps patient data fully private
- Bengali PDF output bridges a real gap in clinical software for the region