I built LOOM because deploying AI models is broken. Every platform needs different conversion (ONNX, TFLite, CoreML), outputs vary between platforms, and Python dependencies make production deployment painful.
LOOM solves this by:
- Loading HuggingFace safetensors directly (no conversion)
- Providing deterministic outputs across all platforms (MAE < 1e-8)
- Compiling to a 10MB binary (vs 2GB+ Python stacks)
- Working on 8 platforms from one model file
Already published to PyPI, npm, and NuGet. First framework with native game engine integration (Godot + LLMs).
The article covers technical architecture, real use cases (healthcare, games, privacy apps), and why local AI matters.
Demos: Desktop (Go server), Godot integration, Android mobile - all linked in article.
Code: github.com/openfluke/loom
Happy to answer questions about the implementation, especially the cross-platform determinism and C-ABI design.
openfluke•2mo ago
I built LOOM because deploying AI models is broken. Every platform needs different conversion (ONNX, TFLite, CoreML), outputs vary between platforms, and Python dependencies make production deployment painful.
LOOM solves this by: - Loading HuggingFace safetensors directly (no conversion) - Providing deterministic outputs across all platforms (MAE < 1e-8) - Compiling to a 10MB binary (vs 2GB+ Python stacks) - Working on 8 platforms from one model file
Already published to PyPI, npm, and NuGet. First framework with native game engine integration (Godot + LLMs).
The article covers technical architecture, real use cases (healthcare, games, privacy apps), and why local AI matters.
Demos: Desktop (Go server), Godot integration, Android mobile - all linked in article.
Code: github.com/openfluke/loom
Happy to answer questions about the implementation, especially the cross-platform determinism and C-ABI design.