Autodidact Software Developer · Generative AI Builder · Audio-Visual Creator

Building local-first and hybrid AI systems where human intent, machine intelligence, software and audiovisual culture meet.

My path into software did not begin in a university or research lab.

It began with questions, experimentation, persistence and the decision to understand how modern computer systems actually work — from operating systems and application architecture to machine learning, generative AI and local inference.

I build software around a simple idea:

AI becomes truly interesting when it stops being just a chat window and becomes a runtime layer inside real software.

My work explores systems that can reason, retrieve information, use tools, process documents, understand speech, generate speech, orchestrate models and interact with local or cloud resources.

I am particularly interested in architectures where cloud AI is optional rather than mandatory.

LOCAL-FIRST AI
     │
     ├── Small Language Models
     ├── Local STT / Speech Recognition
     ├── Local TTS / Voice Synthesis
     ├── ONNX / Edge Inference
     ├── Context + Memory
     └── Tool Execution
              │
              ▼
      AGENTIC APPLICATIONS
              │
     ┌────────┼─────────┐
     ▼        ▼         ▼
    RAG      MCP      Desktop
     │      Tools       Apps
     │        │          │
     └────────┼──────────┘
              ▼
      HYBRID AI ROUTER
       Local ↔ Cloud

My current development work concentrates on:

• Local-first Generative AI
• Small Language Models / SLMs
• Local speech-to-text systems
• Local text-to-speech systems
• Hybrid local + cloud inference
• Agentic AI architectures
• Model Context Protocol / MCP
• Tool calling and agent orchestration
• Retrieval-Augmented Generation / RAG
• Document intelligence
• Long-term context and memory systems
• AI-native desktop applications
• Multimodal and audiovisual interfaces
• Generative visual systems
• OBS / broadcasting automation
• AI-assisted music and media workflows

A large part of my experimentation revolves around keeping the application independent from any single model provider.

flowchart TD

    USER["Human / User"]
    UI["Desktop · Web · Voice UI"]

    USER --> UI

    UI --> ORCHESTRATOR["Agent / Application Orchestrator"]

    ORCHESTRATOR --> MEMORY["Context + Memory"]
    ORCHESTRATOR --> MCP["MCP / Tool Layer"]
    ORCHESTRATOR --> RAG["RAG / Document Intelligence"]
    ORCHESTRATOR --> SPEECH["Speech Layer"]

    SPEECH --> STT["Local STT"]
    SPEECH --> TTS["Local TTS"]

    ORCHESTRATOR --> ROUTER["Inference Router"]

    ROUTER --> LOCAL["Local Models"]
    ROUTER --> CLOUD["Cloud Models"]

    LOCAL --> SLM["SLM / ONNX / llama.cpp / Edge"]
    CLOUD --> AZURE["Azure AI"]
    CLOUD --> OPENAI["OpenAI"]

    MCP --> FILES["Files"]
    MCP --> APIS["APIs"]
    MCP --> SYSTEM["Local System"]
    MCP --> SERVICES["External Services"]

    RAG --> VECTOR["Vector Search"]
    RAG --> DOCUMENTS["Documents / Knowledge"]

The objective is not merely to call an LLM.

The objective is to build an AI runtime capable of deciding where computation should happen and which resources should be used.

I am increasingly interested in running meaningful AI workloads directly on consumer hardware.

Areas I explore include:

Small Language Models
        +
Quantized Models
        +
ONNX Runtime
        +
llama.cpp-style inference
        +
Windows / Edge AI
        +
Local context
        +
Tool execution

The goal is to make AI applications:

• more private
• less dependent on external APIs
• cheaper to operate
• usable offline
• more deterministic
• modular
• easier to integrate into desktop software

Cloud models remain valuable, but they should be a capability, not necessarily the entire architecture.

One of my current areas of development is combining:

Microphone
    ↓
Local STT
    ↓
Local / Hybrid SLM
    ↓
Agent + Tools + Context
    ↓
Local TTS
    ↓
Voice Response

This creates the foundation for AI applications that can communicate naturally without requiring every audio stream or inference request to leave the machine.

Current areas of experimentation include:

• speech recognition
• streaming transcription
• local voice synthesis
• low-latency inference
• voice activity detection
• conversational state
• interruption handling
• local SLM integration
• local/cloud fallback routing

Retrieval-Augmented Generation for Swiss legal information.

JustitiaAI explores AI-assisted access to Swiss legal material using RAG, document retrieval and generative AI.

Core concepts include:

• Azure OpenAI
• Azure AI Search
• vector retrieval
• document ingestion
• source-grounded answers
• citations
• semantic search
• document intelligence
• conversational legal research

The project represents an early step toward more advanced legal AI systems and later experiments such as LexOIMM.

Experimental desktop application combining:

Electron
   +
Vue
   +
Vite
   +
Python
   +
AI interaction

AVA explores the integration of desktop software, Python processing, file workflows, camera input and AI-assisted interaction inside a cross-platform application architecture.

Some of my most active work exists as prototypes, evolving repositories, experiments or private development environments rather than polished public products.

Legal AI · Agentic Desktop System · Document Intelligence

An evolving AI workspace oriented toward Swiss legal research and professional document workflows.

Conceptually, LexOIMM combines:

Electron Desktop Client
        │
        ├── Legal Documents
        ├── PDF Intelligence
        ├── RAG
        ├── Vector Search
        ├── Agent Reasoning
        ├── Persistent Context
        │
        └── Hybrid AI
                 │
          Local ↔ Azure / Cloud

The long-term goal is not simply a legal chatbot, but an AI-native professional workspace.

Experimental generative storytelling and audiovisual system.

Areas include:

• AI-assisted story construction
• generative visual composition
• scene generation
• still-frame pipelines
• visual remix systems
• prompt orchestration
• multimodal generation
• asset libraries
• audiovisual sequencing

The project explores the intersection of:

software × generative AI × visual language × narrative systems

Experimental local conversational runtime combining:

• local speech recognition
• local speech synthesis
• local SLM inference
• contextual memory
• tool execution
• hybrid cloud fallback
• streaming responses

The objective is a lightweight conversational AI layer capable of operating primarily on-device.

Experiments around providing a common orchestration layer between applications and AI systems.

Application
     │
     ▼
Agent Runtime
     │
     ├── MCP Servers
     ├── Open Interpreter
     ├── Local Models
     ├── ONNX Models
     ├── Cloud Models
     ├── REST APIs
     └── Local Tools

A core principle is avoiding tightly coupling an application to one AI provider.

Experimental AI-assisted cryptocurrency research and market-analysis system.

Research areas include:

• portfolio monitoring
• market context
• technical indicators
• news evidence
• on-chain context
• risk analysis
• strategy simulation
• signal scoring
• paper trading
• explainable recommendations

This project is an experimental decision-support and research environment — not a promise of guaranteed financial returns.

A parallel engineering track inside Audioreworkvisions focuses on tools for live audiovisual performance.

Experiments include:

• OBS integrations
• automatic track-ID overlays
• DJ metadata systems
• audiovisual control interfaces
• generative visuals
• media libraries
• video-to-GIF processing
• loop generators
• slideshow engines
• streaming overlays
• live performance utilities

These systems connect my software work directly with my work as an audiovisual creator and Techno DJ.

If a task can run safely and efficiently on the user's machine, local execution should remain a serious option.

Powerful cloud models are valuable, especially for complex reasoning and multimodal workloads.

The architecture should decide when they are required.

Applications should not be designed around one model.

Application Logic
       ≠
Model Provider

Models should be interchangeable components behind a stable application layer.

Useful AI requires more than a prompt.

It requires:

Context
+ Memory
+ Retrieval
+ Tools
+ State
+ Permissions
+ Models

The next generation of AI applications should not only generate text.

They should interact with:

• files
• databases
• APIs
• applications
• operating-system capabilities
• external tools
• structured workflows

My interest is increasingly focused on turning AI concepts into persistent software architectures rather than isolated experiments.

Audioreworkvisions / ARV is where several disciplines intersect:

SOFTWARE
   ×
ARTIFICIAL INTELLIGENCE
   ×
SOUND
   ×
VISUAL SYSTEMS
   ×
TECHNO CULTURE

Alongside software development, I produce audiovisual experiments and Techno transmissions under the Audioreworkvisions identity.

This creates an unusual engineering environment where tools are often tested in real creative workflows rather than artificial demonstrations.

Examples include:

• live visuals
• generative imagery
• DJ interfaces
• OBS systems
• track metadata
• audiovisual automation
• real-time media pipelines
• experimental human-machine interfaces

The creative work and software work are not separate projects.

They continuously influence each other.

My route into technology has been unconventional.

I learn primarily by:

QUESTION
   ↓
RESEARCH
   ↓
BUILD
   ↓
BREAK
   ↓
UNDERSTAND
   ↓
REBUILD
   ↓
CONNECT

I am less interested in memorizing isolated technologies than in understanding how systems connect.

Operating systems lead to application architecture.

Application architecture leads to APIs.

APIs lead to distributed systems.

Distributed systems lead to agents.

Agents lead to models.

Models lead back to hardware.

And eventually everything becomes one system.

My repositories include original applications, experiments, research prototypes, reference implementations and forks used to study technologies across the AI ecosystem.

Current areas of repository research include:

Agentic AI
Local AI
Windows ML
Foundry Local
Azure AI
RAG
MCP
LLM Interfaces
Model Inference
Speech
Desktop Applications
Generative Media

Stay curious & Keep building.

_{Audioreworkvisions · Switzerland}