1.7 KiB
Salka Elmadani
Building Inference-X — better output from the same model.
Universal AI inference engine. Fused computation, adaptive precision, surgical expert loading. 305 KB, 19 backends, zero dependencies. Built in Morocco for the world.
What I build
| Project | What it does |
|---|---|
| Inference-X | Universal inference engine — 305 KB binary, 19 hardware backends, 23 quantization formats, fused dequant+dot kernels, Shannon entropy adaptive precision. Same model, cleaner signal. |
| ATLAS | Open mathematical framework for neural network analysis. |
| Organ Architecture | Neural network surgery — extracting, measuring, and grafting components between AI models to create functional chimeras. |
How it works
The same model produces higher-fidelity output through Inference-X because the computation path is cleaner: fused kernels eliminate intermediate buffers, adaptive precision allocates depth where it matters, and surgical expert loading keeps only active parameters in memory.
A smaller model running through a clean engine can outperform a larger model running through a noisy one.
Philosophy
The best inference engine is the one you do not notice. You should hear the model, not the framework.
Links
inference-x.com · Documentation · Source Code · Elmadani.SALKA@proton.me
Morocco