What I've built

I designed and built the association platform: public site, members area and a custom CMS for admins. New members enter a private area where they manage their profile, dues and get access to all digital member services.

On access I went beyond the minimum: one-time codes, an error log, and visibility rules written in the database. Every morning an automated process automatically notifies anyone who crosses dues waiver thresholds.

A national association that used to live on spreadsheets and email: members now sign up online, pay dues, find each other and run mentorship inside the same platform.

I built more than ten internal tools for people who had never written a prompt, each cut to fit existing workflows.

There are apps that estimate revenue trends, others that analyze period results and say where you are making and losing money, others that produce reporting, others that run quality checks on work done by other AIs.

I always worked inside a single model ecosystem, Google Gemini, both in the individual version and on cloud infrastructure.

I designed an agentic AI pipeline that every day finds local businesses with outdated sites, visits their pages, analyzes them and prepares the ground for a commercial proposal. The analysis covers text and visuals and assigns an obsolescence score.

Every contact moves through several states, from discovery to conversation. Each agent has a model suited to its task, with a fallback if the first does not respond, and every call is logged with its price tracked.

Site generation stays the human-in-the-loop point: the pipeline prepares everything, the human decides and launches. After send the system continues on its own, reads replies, sends follow-ups, archives for GDPR.

I designed and ran an internal training course on applied AI. Every session started from a daily difficulty: how to build a presentation the right way, how to analyze tickets, how to build an automation, how to make a browser extension that does the thing no commercial product does, and so on.

For each session I prepared materials and a set of ready-to-use instructions, which I later shared as a prompt library.

This is the only training activity that sits here next to what I built because I am particularly attached to it: it was the first time I generated AI training from scratch, and it is the one I care about most.

You're inside this project, the platform you're using right now. Every morning an automated process reads industry sources, picks what matters and prepares the AI news roundup; on Sunday a long article comes out that I publish after review.

I spread requests across multiple providers, picking the right model for each task, and put a counter on top that logs every single call with its cost and stops on its own when daily spend crosses the threshold I set.

On the same foundation run the public tools lab, the book members area, the newsletter, the admin console with two-factor access and the whole site in two languages.

I connected bank accounts from 3 banks so they update automatically every day, and built a two-level classification: first the rules I wrote, then a language model that steps in only where the rules fall short.

In the backend I built recognition of internal transfers, automatic credit of cash withdrawn at the ATM, and recurring forecasts that latch onto the real movement when it arrives.

On access I did not cut corners: password, one-time code, device recognition and an email alert when something looks off.

I designed a digital area around my book that continues where paper stops. Each chapter unlocks with a word hidden in the pages, and anyone who collects them all reaches a thirteenth secret chapter that does not exist in the book.

Inside there are ready-to-copy prompts, operational checklists, external resources curated chapter by chapter and a set of guided generators that build the right prompt by answering a few questions. I added a per-chapter changelog so readers see what is new at a glance, and a channel for readers to suggest new resources.

I built a web app that creates a guided path bringing sales to a motivated decision, instead of an impression when qualifying prospects. Twenty branched questions: each answer opens the next piece and closes what is no longer needed, with the option to go back without losing what was already entered.

The path has five checks with related alerts for possible issues on number sustainability. Behind it runs a calculator that updates the margin on each conversion with every figure.

At the end the model rereads everything collected and runs an AI analysis: whether the contact makes sense as-is, or which terms need renegotiating for the job to make sense.

I built two extensions that bring AI into heavily used sites. For LinkedIn I created seven features to improve the experience: a score that rates posts, a check that dismantles inflated posts from others, a booster to show more profile info, and so on.

On YouTube I used the model's visual side: for example the extension captures the frame I'm watching and explains it or opens a conversation about the video content.

Both work with the user's API key saved only in their browser: requests go straight to the provider, I have no server in between and I see nothing of what passes through.

I built ClauDisk, a Chrome extension that keeps a living mirror of Claude.ai projects in a folder on disk. It is not a one-shot export: while you browse Claude, chats, docs, memory and artifacts land as Markdown ready for Cursor or Claude Code.

Sync starts in dry-run: you see what would be written and confirm only when you want. Soft-delete, attachment size caps, vault verify and archive stats stay local. No ClauDisk server in the middle: everything goes through your browser session and the folder you choose.

It is open source, MIT, with English UI by default and Italian in Settings. It runs on Chrome and Edge; the product site covers install and the flow.

That question led to GEO Visibility Tracker: it queries four LLMs in parallel on the same prompt and classifies each answer by whether the brand is recommended, only mentioned or ignored entirely. Once that was built, I realized the pattern held for more, and three others followed.

CV Dashboard reads a resume, compares it to a job post, analyzes and produces 6 different cover letters. Google Ads Genius generates all the data needed for Google Ads. Travel Planner AI builds a detailed day-by-day travel itinerary from a few inputs.

All four share the same architectural choice: the user brings the key, the request leaves their browser and goes straight to the provider without passing through my servers.

I built an application that, in a room of strangers, surfaces the connections nobody would have found alone.

Each participant fills in their profile on their phone in under a minute: role, sector, strength, what they are looking for. Profiles land in real time on the dashboard projected in the room, and when the group is complete I run the matching: one model call that reads every profile together instead of comparing them in pairs, and matches people by their interests.

Everything lives in a single file, with no framework and no installed dependencies.

I built a tool to facilitate any brainstorming workshop where the host opens a room and participants join from the browser with a name and a short secret code, with no signup and nothing to install. From there everyone works in sync on the same board, and every change propagates live.

I designed the path as seven timed phases, from warm-up to final decisions, through brainstorming, point-voting rounds and an impact-effort matrix where you drag the top ideas.

The host stays in control: they advance phases, run the shared timer and can lock the board when needed. At the end the system processes the results and generates a final AI recap.

I built an app for yearly work planning that tracks holidays, leave, business trips, sick leave and days from home, and warns me when I'm about to exceed weekly or monthly limits.

It calculates on its own, there is a leave-time simulator, it works offline, installs as a real app and has no server behind it: data stays in the browser, privacy-first.