Aurelio Capriello AI Conductor ENIT
Research & experiments

Beyond client work, I run deep technical experiments — and I publish the ones that fail, not just the ones that work. A method that only reports its wins is marketing, not engineering.

AI-native OS kernel

A from-scratch operating-system kernel in Rust with zero external crates — boots via UEFI to a graphical shell in ~195 KB, with an in-kernel AI inference experiment. Research into what an AI-native OS could be.

⚠ Boots in emulation and early hardware; full hardware bring-up (PCI, USB-HID) is unfinished. Research, not a shipped OS.

Frontier models on consumer hardware

Local LLM inference at scale: exact 7B inference in 2.2 GB VRAM, and JIT expert-loading for Mixture-of-Experts that exploits temporal locality — the path to running big models on the machine you already own.

⚠ Honest result: the locality is real but moderate — my own pre-registered PASS criterion FAILED on a load-balanced MoE. Published anyway; the negative result is the finding.

What makes a model generalize

~140 GPU experiments on the real question: does a model understand, or interpolate? Finding: structure beats raw parameters out-of-distribution — with the exact boundaries measured, and several of my own predictions falsified along the way.

⚠ Exploratory research. Notes and reproducible scripts available on request.

The graveyard of falsified ideas is documented, on purpose. That discipline is exactly what I bring to a client's system.

Contact

Have an AI problem that needs to actually work?

From a one-hour consultation to a complete build: AI adoption, custom software, assistants, automation, reliability engineering, security audits. Everything verifiable is on GitHub; the rest I'll show you live. Remote worldwide, based near Milan.