Why UnaMentis?
Existing voice AI tools struggle with extended educational sessions. UnaMentis was built from the ground up to enable natural, flowing conversations that can last 60-90+ minutes without degradation.
Voice-Native Design
Built for voice from day one. Natural interruption handling, turn-taking logic, and voice activity detection create fluid conversations.
Provider Agnostic
Swap STT, TTS, and LLM providers without code changes. Support for OpenAI, Anthropic, ElevenLabs, Deepgram, and self-hosted options.
Curriculum System
UMLCF (UnaMentis Markup Language Curriculum Format) is a purpose-built specification for conversational AI tutoring with voice-optimized content.
Low Latency
Sub-500ms median end-to-end turn latency through careful architecture, prefetching, and intelligent routing.
Self-Hosted Ready
Run your own STT, TTS, and LLM servers for privacy-first deployments. Full support for Ollama, llama.cpp, Piper, and more.
Full Observability
Real-time telemetry, cost tracking per provider, latency metrics, and thermal monitoring for production deployments.
Architecture Overview
Supported Providers
UnaMentis supports a wide range of providers for each component of the voice pipeline. Mix and match based on your needs for quality, cost, latency, or privacy.
Speech-to-Text
- Cloud AssemblyAI Universal
- Cloud Deepgram Nova-3
- Cloud OpenAI Whisper
- Device Apple Speech
- Device GLM-ASR-Nano
- Self-Hosted GLM-ASR Server
Text-to-Speech
- Cloud ElevenLabs Flash/Turbo
- Cloud Deepgram Aura-2
- Device Apple TTS
- Self-Hosted Piper
- Self-Hosted VibeVoice
Large Language Models
- Cloud OpenAI GPT-4o / 4o-mini
- Cloud Anthropic Claude 3.5
- Self-Hosted Ollama
- Self-Hosted llama.cpp
- Self-Hosted vLLM
- Device llama.cpp (experimental)
Voice Activity Detection
- Device Silero VAD (CoreML)
- Device TEN VAD
- Device WebRTC VAD
UMLCF Curriculum Format
UMLCF (UnaMentis Markup Language Curriculum Format) is a JSON-based curriculum specification designed specifically for conversational AI tutoring.
Voice-Native
Every text field has optional spoken variants optimized for TTS pronunciation.
Standards-Grounded
Maps to IEEE LOM, LRMI, SCORM, xAPI, QTI, CASE, and Open Badges.
Tutoring-First
Built for natural conversation with stopping points and misconception handling.
Unlimited Depth
Topics can nest to arbitrary depth for complex subject matter.
{
"umlcf_version": "1.0.0",
"curriculum": {
"id": "intro-machine-learning",
"title": "Introduction to Machine Learning",
"topics": [{
"id": "supervised-learning",
"title": "Supervised Learning",
"learningObjectives": [
"Explain the difference between supervised and unsupervised learning",
"Identify when to use classification vs regression"
],
"transcript": {
"text": "Let's start with supervised learning...",
"spokenText": "Let's start with supervised learning.",
"stoppingPoints": [{
"afterParagraph": 2,
"comprehensionCheck": "Can you give me an example of a labeled dataset?"
}]
}
}]
}
}
Server Components
UnaMentis includes a Python-based management server and web dashboard for monitoring, curriculum management, and analytics.
Management Server
Async HTTP server with WebSocket support for:
- Remote logging aggregation from iOS clients
- Real-time metrics streaming
- Resource monitoring (CPU, memory, thermal)
- Idle state management
Python 3.11+ / aiohttp / asyncio
Curriculum Database
Storage and retrieval for UMLCF curricula:
- File-based storage for development
- PostgreSQL support for production
- Search and filtering by metadata
- Topic hierarchy navigation
PostgreSQL / File-based / JSON
Web Dashboard
React-based administration interface:
- Curriculum browsing and management
- Session analytics visualization
- Real-time metrics streaming
- Provider health monitoring
Next.js / React / TypeScript
Technical Stack
Mobile Client (iOS)
- Swift 6.0 with strict concurrency
- SwiftUI for all views
- AVFoundation for audio
- CoreML for on-device ML
- Core Data for persistence
Backend
- Python 3.11+
- aiohttp for async HTTP
- PostgreSQL (production)
- Next.js dashboard
- WebSocket for real-time
Performance
- <500ms median turn latency
- <1000ms P99 latency
- 90+ minute session stability
- <50MB memory growth
- Thermal monitoring
Quality
- 119+ automated tests
- Real-over-mock testing
- SwiftLint / SwiftFormat
- Accessibility-first UI
- Comprehensive docs
Team
UnaMentis is developed and maintained by a dedicated team committed to making high-quality AI tutoring accessible to everyone.
Richard Amerman
Founder and Project Lead
Cy Goerdt
Partner
Ready to Get Started?
UnaMentis is fully open source and ready for contribution. Whether you want to use it, extend it, or help improve it, we welcome you.