Nicolas Schlaepfer is an AI engineer developing decentralized AI inference, editable agent workflows, and models optimized for consumer devices. His work spans peer-to-peer language-model infrastructure and Apple-native adaptations of text-generation and neural-signal systems.
At Hyperspace, Schlaepfer worked on an AI OS application that used llama.cpp to distribute inference across personal computers running Windows and macOS. A Mistral model he developed, funded by Hyperspace AI, used a custom dataset created with GPT-4-assisted automation.
At AI Engineer World’s Fair 2024, Schlaepfer demonstrated an agent architecture built around three practical ideas:
- Editable agent workflows: HyperEngine V3 translated requests into JSON-encoded directed acyclic graphs and streamed tasks into a React Flow editor, allowing users to change objectives, expected outputs, and individual steps.
- Specialized model orchestration: Qwen2 Instruct handled reasoning while Llama 3 70B produced summaries. Tasks incorporated the overall objective, preceding results, web content converted into Markdown, and generated Python executable through an integrated terminal.
- User-directed agent infrastructure: A virtual file system, visual planning interface, code execution, and prospective memory capabilities established reusable components for agents whose behavior remained inspectable and adjustable.
Schlaepfer subsequently published Core ML adaptations through his oraculumai profile. His Apple-native LLaDA implementation exports a single denoising pass while leaving iterative diffusion sampling to the application. His ZUNA Core ML conversion adapts an existing EEG reconstruction model for Apple platforms, supporting signal denoising and missing-channel reconstruction; his contribution is deployment adaptation, not authorship of the upstream model.
