Hi there! I'm Joe.

I build software and measure AI on hardware I own, then publish what actually happened, run logs included.

CAIO [Local, Enterprise] by day. After hours it’s the home lab: one RTX 5090 & EVGA 3090 ULTRA, vLLM and llama.cpp, and a growing pile of run logs. This site is the public notebook for both.

A few things about me

The rig

NVIDIA GeForce RTX 5090 (32 GB), AMD Ryzen 9 9950X3D, 64 GB DDR5. One box, one GPU, no cloud credits.

The bench model is for the 3090, goal is test how Ampere quants hold up against Blackwell, hoping to lower barrier to entry.

Everything on the bench page was measured on it.

Numbers over vibes

Every benchmark here ships with its methodology and run log. Determinism is checked, not assumed: identical runs have to produce identical output before a result counts.

Day job

CAIO [Full Stack = Local AI in prod].

After hours

Squeezing more out of the 5090/3090: quantizing a 27B hybrid model until a whole novel fits in one prompt, and writing up the four walls that stood in the way. The wins and the regressions both get published.

Shipped things

The projects page is the portfolio: live links, source, and the story of how each one was built.

Say hi

Email is the best way to reach me. Twitter is another great place, find me @JoeGuglielmucci

Now

By the numbers

1,594 commits in the past year across every project, deduped by hash.

Weekly activity

Commits per week across every local project, trailing 12 months.

0 60 120 OctNovDecJanFebMarAprMayJunJulAugSep 118, best week

source: git history, deduped by commit hash · computed 2026-09-23

Language share

Lines of tracked source across active projects, by volume.

JavaScript: 58%TypeScript: 25%Markdown: 8%SQL: 4%Other: 5%
  • JavaScript58%
  • TypeScript25%
  • Markdown8%
  • SQL4%
  • Other5%

source: local repositories, tracked source only · computed 2026-09-23