Joe Fioti

Joe Fioti is the co-founder and chief executive of Luminal, an AI infrastructure company developing compilers and inference systems that run machine-learning models efficiently across different hardware. Previously an Intel engineer working on CPU microcode and machine-learning accelerators, he builds software that translates increasingly complex models into programs tailored to individual chips.

Fioti founded Luminal with Matthew Gunton, formerly of Amazon, and Jake Stevens, formerly of Apple. The company joined Y Combinator’s Summer 2025 batch and raised $5.3 million in a seed round led by Felicis that November.

His open-source machine-learning compiler represents neural networks through a small collection of primitive tensor operations and static computation graphs. Matrix multiplication and convolution emerge from combinations of simpler arithmetic, reductions, and tensor-shape transformations. Most transformer computation can be specified in advance, with limited flexibility retained for changing sequence lengths and key-value caches.

• Search-based deep-learning compilation. Fioti uses simple rewrite rules to generate mathematically equivalent GPU kernels, then searches for efficient implementations. Luminal has automatically discovered a FlashAttention-style optimization without encoding that complete algorithm as a special case.

• Whole-graph optimization and megakernels. A complete model graph allows Luminal to fuse operations, reuse memory buffers, dispatch kernels ahead of time, and reduce expensive data movement. Fioti’s work on megakernel compilation extends these techniques to deeper execution pipelines and finer-grained synchronization. An external automatic-differentiation extension applies the same compiler optimizations to training graphs.

• Hardware-aware heterogeneous inference. Luminal can divide inference between different accelerators: its collaboration with Positron pairs GPU-based prompt processing with token generation on Positron’s Atlas hardware. Fioti has also tracked cloud GPU availability and pricing to match customer demand with available compute.

• Numerical fidelity in model serving. Fioti argues that inference optimizations must preserve output quality alongside speed and cost: aggressive quantization and reduced-precision arithmetic can accumulate errors that undermine otherwise attractive performance gains.

Luminal’s Inference OS extends this compiler work into managed and on-premises deployment across CPUs, GPUs, and specialized accelerators.