VibeUI
A small design system built for AI products specifically: streaming text, thinking states, model pickers, the stuff every AI app ends up rebuilding from scratch. Concept components, ready to steal from.
DESIGN × CODE - Apps, Products & AI
This is where I build things. I design apps, then code them until they actually run on a phone. AI products too, and way too many prompts. Some of it ships. Some of it breaks. All of it teaches me something. Not a portfolio. A laboratory.
booting gianluca.os...
[✓] curiosity module loaded
[✓] app build: passing
[✓] ai toolkit mounted
[✓] design sense: online
[✓] caffeine: critical
ready.▋
A small design system built for AI products specifically: streaming text, thinking states, model pickers, the stuff every AI app ends up rebuilding from scratch. Concept components, ready to steal from.
A pattern library for how interfaces could actually show their confidence, not just a chat bubble that sounds sure of everything. Citations, uncertainty markers, and 'I might be wrong' states, explored as UI.
A workbench for saving, tagging, and re-testing the prompts that actually work, instead of losing them in fifty scattered chat tabs. Still just a concept, but it's the tool I keep wishing existed.
Most AI products dump you into a blank chat box and hope for the best. A redesign concept for the first five minutes: showing what the tool can actually do before asking you to guess.
A concept tool for the way I actually compare AI outputs: two responses side by side, scoring criteria built in, notes attached to the ones that need a second look. Grew out of doing this same thing in a spreadsheet for months.
Most rubrics score correctness. This one tries to score honesty. Does the model hedge when it should, or sound confident about something it's guessing? Concept work, grown out of months of rating real responses.
Two years rating AI responses, one rubric at a time. Some days it's catching a hallucination in two seconds flat, other days it's untangling something messy across text, images, even code, and still hitting the deadline. Hundreds of responses in by the end, and a pretty good nose for when a model is bluffing.
Two AI voices, same prompt, and I was the one deciding which one actually nailed it. Pronunciation, how close it stuck to the original ask, whether the audio has weird glitches, how natural it sounds instead of robotic. Side by side, over and over, until the better one is obvious.
Personal Project
My own take on Jarvis. Still in progress, but it's the project I keep coming back to.
Building FORGE
> what it is
An AI assistant built to actually know your context instead of starting from zero every time you open it. Claude Code runs the thinking, Obsidian holds the memory: notes, decisions, everything it needs to remember so you don't have to repeat yourself.
> status
in progress, long-term build. no demo yet, no public repo.
> built with
~/forge ▋