AI Stack

Hardware

• GPU:NVIDIA RTX 4000 Ada (20 GB VRAM)
• Modelle:devstral-small-2:24b (15GB), qwen2.5-coder:14b, more

Components

Ollama (LLM Runtime)

• URL:http://ollama:11434(internal)
• GPU:Exclusive, container withnvidiaRuntime
• Modelle:Automatically loaded via Open WebUI

Open WebUI (Chat Interface)

• URL:https://ai.xynap.tech(auth-cut)
• Features:Multi-Model Chat, RAG, Prompt templates, API keys

Whisper (Speech-to-Text)

• URL:http://whisper:9099(internal)
• Modell:whisper-large-v3-turbo
• Endpoint: POST /asr?output=json

Piper TTS (Text-to-Speech)

Two services:

Service	Port	Protocol	Use
piper-tts	10200	Wyoming	FreeSwitch
piper-http	5100	HTTP REST	Web apps

German voice:thorsten_emotional(medium quality)

LibreTranslate

• URL:http://libretranslate:5000(internal)
• Sprachen:DE, EN, FR, ES, IT, PT, ...
• API: POST /translate {"q": "text", "source": "de", "target": "en"}

AI Agent (Genesis)

The AI Coding Agent uses:

• Ollama LLM for code generation
• Redis PubSub for Task Communication
• Platform API as control interface

# Task erstellen
POST /api/v1/coder/tasks
{"prompt": "...", "workspace": "/path/to/code"}

# Status abrufen
GET /api/v1/coder/tasks/{token}

Real-time interpreter

AI-based telephone interpreter:

Anruf → FreeSwitch → Whisper (STT) → LibreTranslate → Piper (TTS) → zurueck

Substanced: DE└EN, DE└TR, DE└AR and other language pairs.