clivora
AI-powered medical transcription platform
A full-stack SaaS platform that turns spoken medical consultations into structured clinical reports — built from scratch with custom AI pipelines, subscription management, and multi-language support for the Central European healthcare market.
Record a consultation. Get a structured clinical report. No typing required.
Doctors spend 2+ hours a day typing notes instead of treating patients.
Healthcare professionals are legally required to document every consultation. In Hungary and Central Europe, no AI transcription tool supported Hungarian medical terminology. Generic speech-to-text tools failed on clinical language, produced unusable output, and had no integration with medical workflows. Physicians were stuck with manual typing or copy-pasting from previous notes — 15 to 30 minutes per patient, every day.
Existing solutions like Dragon Medical were US-centric, English-only, and prohibitively expensive. The Central European market had zero viable AI-powered alternatives. The gap was clear: a platform that could accurately transcribe Hungarian medical speech, structure output into proper clinical formats, and support multiple specialties — all within a browser.
three-click workflow
Select a specialty. Hit record. Get a structured report. The entire platform was designed around eliminating friction for busy doctors who are resistant to new tools and have no time for learning curves.
Language-first
Hungarian medical transcription required custom-trained models. Off-the-shelf solutions achieved ~70% accuracy on medical terminology — unusable for clinical documentation. We built a hybrid AI pipeline from scratch.
Template-driven output
Doctors don't want raw transcriptions. They need structured reports — SOAP notes, discharge summaries, consultation reports — matching their specialty's documentation standards. 20+ medical domains, each with tailored templates.
Self-hosted AI
Using open-source models (Whisper + GPT OSS 120B) instead of commercial APIs gives cost control, data sovereignty, and the ability to fine-tune for Hungarian. Healthcare data stays on our infrastructure — critical for GDPR.
Separated architecture
The ML inference backend scales independently on AWS AppRunner. The Laravel SaaS platform and React transcription app communicate via JWT/SSO. Each tier scales, deploys, and fails independently.
the transcription pipeline
Not a wrapper around an API. A three-stage custom pipeline combining fine-tuned speech recognition with large language model refinement — built specifically for Hungarian medical language.
Raw transcription
Browser-based audio capture using a custom PCM AudioWorklet running in a separate thread. 44.1kHz recording, automatic resampling to 16kHz client-side (60%+ upload reduction), silence detection, and configurable snippet chunking with overlap.
Audio streams to a fine-tuned Whisper model that handles Hungarian medical vocabulary with 95%+ accuracy — up from ~70% with the untuned base model.
AI refinement
Raw transcription passes through GPT OSS 120B for medical context understanding. The LLM corrects terminology errors, adds proper punctuation, resolves ambiguous medical abbreviations, and structures the text into coherent clinical language.
This hybrid approach — combining ASR with LLM post-processing — produces results far superior to either model alone.
Structured output
Refined text is mapped to specialty-specific templates. A gastroenterology consultation produces different sections than a cardiology follow-up. 100+ pre-built medical snippets compose into formal reports matching each domain's documentation standards.
The final report is editable in a rich text editor, ready to copy into any EHR system.
real-time capture
Audio capture runs entirely in the browser — no plugins, no installs. A custom PCM AudioWorklet processor operates in a separate thread to prevent UI jank during recording. Automatic silence detection and configurable snippet duration handle continuous consultations of any length.
Client-side resampling from 44.1kHz to 16kHz reduces upload payload by 60%+ before transmission. Base64 encoding and streaming upload ensure reliable delivery even on unstable hospital networks.
specialty-aware output
20+ medical domains — gastroenterology, cardiology, surgery, general practice, and more. Each specialty has tailored templates composed of reusable medical snippets that match real clinical documentation standards.
Templates are language-aware across Hungarian, English, German, and French. Doctors configure their preferences once — specialty, language, preferred report format — and every subsequent transcription follows their setup automatically.
full SaaS architecture
Not just a transcription tool. A complete three-component production system — each built custom, each scaling independently.
JWT/SSO cross-app auth
RS256 JWT tokens with 15-minute access TTL and 30-day refresh tokens bridge the Laravel platform with the React transcription app. The React app never sees the private key. JTI-based refresh tokens enable O(1) database lookups.
Single-use SSO login codes with 5-minute expiration. Recording protection delays auth redirects during active sessions — a doctor mid-consultation never gets kicked out by a token refresh.
Stripe billing engine
Subscription plans with usage-based limits, a trial credit system (5 free transcriptions), and webhook-driven lifecycle management. Every transcription call validates subscription status and checks usage limits with under 200ms overhead.
Layered middleware: JWT decode → user resolution → subscription check → usage counter → upstream proxy. All with early-exit patterns to minimize latency on every API call.
Real-time text display
A custom React hook (useStreamingText) buffers incoming transcription text and renders it character-by-character at 30fps. Intelligent cursor preservation and user-scroll-detection prevent fighting the doctor's reading position.
30-frame buffering ensures smooth animation even with irregular network delivery. The editor remains fully interactive during streaming — doctors can start editing before transcription completes.
AWS auto-scaling
The GPU-intensive ML inference backend deploys on AWS AppRunner, auto-scaling independently of the web application. Transcription uploads proxy through Laravel to the inference service, allowing horizontal scaling of each tier.
Docker containers, pnpm monorepo with shared packages, AWS CodeBuild for CI/CD, S3 for encrypted storage. Database-backed sessions and cache enable multiple web server instances.
administration
14 fully-featured Filament admin resources covering every aspect of the platform — from user management to medical data curation to revenue analytics.
Account management
User CRUD with role-based access control (Spatie Permissions), admin impersonation, subscription lifecycle management, and usage tracking per account.
- User accounts with role management
- Subscription plan assignment & overrides
- Usage history & limit enforcement
- Admin impersonation for support
Templates & snippets
Medical domain management, specialty-specific template editing, reusable snippet library, and language variant configuration across all four supported languages.
- 20+ medical specialty domains
- Template builder with section presets
- 100+ reusable medical snippets
- Multi-language variant editing
Revenue & engagement
Dashboard widgets for MRR tracking, churn rate, user engagement metrics, transcription volume, and trial-to-paid conversion rates. All data sourced from Stripe webhooks and internal usage logs.
- MRR & revenue charts
- Churn rate monitoring
- User engagement & retention
- Trial conversion tracking
Hungarian medical transcription accuracy
Recording to structured report
Medical specialties supported
Documentation time reduction
Healthcare data demands healthcare-grade security.
Data sovereignty
All data stored in EU region on AWS infrastructure. Self-hosted AI models mean patient audio and transcriptions never leave controlled infrastructure. Encrypted storage on S3 with strict access controls and audit trails via JSON payload storage.
Authentication & compliance
RS256 asymmetric JWT tokens, hashed refresh tokens with IP and user-agent tracking, single-use SSO codes, Stripe webhook signature verification, CAPTCHA on public forms. GDPR compliant with HIPAA-aligned data handling practices.
Custom AI pipeline
Not a wrapper around a commercial API. Fine-tuned Whisper + GPT OSS 120B post-processing — built, trained, and hosted specifically for Hungarian medical language. 95%+ accuracy where off-the-shelf achieved 70%.
Full-stack ownership
One developer owning the entire stack — Laravel SaaS, React transcription app, ML backend, AWS infrastructure, Stripe integration, admin panel. No handoff gaps, no integration seams.
Domain-first design
Every feature shaped by physician workflow research. Three-click operation, specialty-aware templates, recording protection during auth flows. The tool adapts to doctors — not the other way around.
Production-grade infrastructure
Independent auto-scaling tiers, JWT-based service mesh, encrypted storage, audit trails, GDPR compliance. Built for healthcare reliability from day one — not retrofitted later.
From 30 minutes per patient
to under 2.
Clivora is the first AI-powered medical transcription platform supporting Hungarian language — filling a gap that left Central European physicians without any viable alternative. Documentation time dropped from 15–30 minutes per patient to under 2 minutes: record, review, edit, done.
The platform serves 20+ medical specialties with structured report templates, supports four languages for expansion across the Central European market, and operates as a sustainable subscription-based SaaS with Stripe-powered billing and trial-to-paid conversion tracking. Active development continues with EHR integration and expanded specialty coverage.
have an idea?
let's make it work.
Send us a message and we'll get back to you quickly.