MedDiktat
Matija Žiberna
Like this project
Posted Jan 28, 2025
Built MedDiktat—a medical transcription tool for Slovenian doctors. Engineered a a system using AI model, fine-tuned with 1,000+ transcription pairs
AI-powered medical transcription tool aiming to assist doctors at private clinics to transcribe their medical examination and issue reports faster
🚀 Project Overview
Problem: Slovenian doctors waste hours manually transcribing patient notes, risking errors and burnout.
Solution: Built a secure, AI-powered tool to automate medical dictation transcription while preserving doctor-specific formatting preferences.
Key Features:
Slovenian speech-to-text fine-tuned for medical terminology
Real-time transcription progress via WebSockets
GDPR-compliant patient data workflows
Custom formatting presets per doctor
Browser-based recording with auto-save fallback
My Role: Sole Full-Stack Developer & Product Designer
Timeline: 3 months (MVP to launch)
🛠️ Technical Stack
Frontend:
Next.js (App Router) + Vite (React)
Shadcn UI + Tailwind
WebSockets for real-time updates
Backend:
Node.js 23 + Express + Typescript
PostgreSQL + Prisma ORM
Google Cloud Tasks (queue management)
AI/ML:
Google Vertex AI Speech-to-Text API
Custom model trained on 1,000+ doctor-corrected Slovenian transcripts
DevOps:
Dockerized monorepo (FE/BE/DB)
Hetzner VPS (Ubuntu 24) + Nginx
GCP Storage (temp) + Backblaze (archive)
Auth: Auth0 (JWT sessions + user caching)
💡 Technical Challenges & Solutions
1. Slovenian Medical Speech Recognition
Trained Vertex AI model using paired raw/corrected transcripts
Added context-aware correction (e.g., "kardio-log" vs "kardiolog")
Result: 95% accuracy for specialist terminology
2. Real-Time Progress Tracking
Implemented WebSocket pipeline:
Client → Cloud Task → Transcription API → Socket.IO → UIMocked 50% progress on task queuing to reduce perceived latency
3. HIPAA-like Compliance
End-to-end encryption for audio/files
Auto-purge temp GCP storage after processing
🔗 Links & Documentation
Live Demo: meddiktat.si
🎓 Learning Outcomes
Technical:
Optimizing Google Speech-to-Text for low-resource languages
WebSocket-backed progress systems
Cost-effective hybrid cloud storage (GCP + Backblaze)
Business:
Doctors prioritize workflow preservation over "smart" features
Stakeholder education for AI accuracy expectations
💼 Business Overview
Target Audience:
Slovenian private practitioners (orthopedics focus)
Specialists managing 20+ daily patient notes
Market Validation:
Pilot group of 2 physicians during beta testing