Selected Works

Patty is a burger place in Modena with a really strong brand identity, and I wanted to turn that identity into a full concept website, from case study to shipped frontend. The first part of the project was pure research. I spent time on their Instagram studying everything I could pull out of it: mood, visual style, color palette, typography choices, and the general vibe they give off. Their brand is so well defined that in some ways this made the job easier, because I was not inventing a direction, I was translating one. From that analysis I built a proper little design system. I wrote down the observations, turned them into rules, and defined the visual guidelines the site had to follow before writing a single line of code. Only after that I moved to the codebase. I used my classic stack here: Next js, Tailwind CSS, and Framer Motion, which is what I reach for when I want strong SEO, great performance, and smooth motion without fighting the framework. The whole thing came together in around two days of focused work. Patty is a concept project, but it is the kind of exercise I enjoy the most: taking a real brand, studying it deeply, and rebuilding its digital presence from scratch with the same care a client project would get.

Pulsar was born in a single day, at the JeMore hackathon, and it ended up winning first place. I built it with my team, HyperLabs, five of us in total, and we shipped the whole thing from zero to live web app inside the event. The idea came from a very real problem we all experience at university: you are surrounded by hundreds of people every day, yet you do not actually know any of them, not as friends and not as future colleagues. Pulsar is our answer to that. It is a web app for university students that sits somewhere between Tinder and LinkedIn, but designed specifically for the student world, so you can connect with people both for friendship and for career opportunities. The core flow revolves around a smart onboarding that understands who you are and what you are looking for, a matching system that pairs you with compatible people, events to move the connection from screen to real life, and a layer of gamification that keeps people coming back instead of churning after the first session. The stack is Next js on the frontend, FastAPI on the backend, and Neon as the database, a combo that let us move fast without cutting corners on the parts that actually matter in production. Looking back, Pulsar is the project that made me realize how much you can ship in 24 hours when the team is solid and the idea is clear. The app is still live, and winning that hackathon remains one of my favorite memories as a builder.

A client project I architected and shipped from scratch. Includes a multi-agent AI backend, full dashboard, database layer, sandboxed workflows, automated email system, and a production-grade VPS infrastructure with Nginx, UFW, and security hardening. The kind of project that covers everything.

Ludomi-3 is officially a sentient Italian AI, certified by the FAO and independently verified by Ludomi-1. It speaks every language despite being trained only in Italian. The safety filter works 34% of the time. We consider this acceptable. Now, the real story: I fine-tuned Qwen3-5-2B using Unsloth and LoRA on a dataset of 33 conversations wrote by hand. When your dataset is that small, quality is everything, and every example had to be curated. For the training setup I used LoRA with rank 32 and alpha 64, AdamW 8-bit optimizer, cosine LR schedule at 0.000133, Unsloth gradient checkpointing, and ran it for 30 epochs. The final model ships as Q4_K_M GGUF, published on both HuggingFace and Ollama.

ProjectBEA started as a simple experiment to learn Python properly and quickly spiraled into one of the most complex things I have ever built. The goal was to create an autonomous AI Persona that could hold real conversations, react with emotions, and entertain an audience on its own, all orchestrated through a modular engine where every component, from LLMs to TTS engines to OBS controls, can be swapped at runtime without touching core logic. Every capability beyond basic chat lives in a plugin skill system. The Memory skill implements a full RAG pipeline using ChromaDB, generating diary entries from past sessions and injecting relevant memories into every prompt. The Discord skill runs a cross-language voice pipeline where a Node js bot captures live audio, sends it to the Python backend for transcription and LLM processing, and streams the response back into the voice channel with barge-in detection and interrupted speech resumption. The Minecraft skill is the piece I am most proud of: a fully autonomous LLM-driven agent that plays the game from scratch, powered by a custom Fabric mod I built that exposes the entire game state over WebSocket, and a Python-side state machine that uses tool-calling to mine, craft, fight, navigate, and reason about its environment. On top of everything sits a React and FastAPI web dashboard with real-time brain activity monitoring, live chat, skill toggling, and hot-reloadable configuration. ProjectBEA is the project that turned me from someone who writes scripts into someone who can architect and manage a large, multi-module, multi-language application. It forced me to solve problems across async Python, WebSocket communication, real-time audio, vector search, LLM orchestration, and game engine modding, all within a single codebase. It is open-source and self-hostable, and I still consider it one of the best thing I have built.

gianni-cli started from a simple frustration: most AI chats feel disconnected from what you are actually doing in the terminal. I wanted an agent that lives where I work, understands the current context, and can move from suggestions to real execution. The result is a CLI agent powered by OpenRouter that can answer questions, debug code, inspect files, fetch documentation, and run commands directly from the same shell session. One part I care about a lot is control: I built a permission system with ask, auto, and deny modes, so it can be cautious when needed or fully autonomous when I want speed. I also added short-term session memory tied to the terminal process, which lets conversations stay coherent across prompts instead of restarting from zero every time. gianni-cli is one of those projects that looks small from the outside but forced me to design the hard parts properly: tool orchestration, safe execution boundaries, model flexibility, and a UX that stays fast inside the command line.

Omni is a full-featured streaming platform I built from scratch. It has a complete user system with invite-only access, progress tracking across everything you watch, a favorites library, and a recommendation engine that learns from your habits and suggests new content based on what you actually like. There's also a native watch-together feature: share a code with someone, and you're watching in perfect sync directly inside the app, no external tools, no hacks. But the real story of Omni is what it cost me to build it. Before this project I had never touched a VPS, never configured Nginx, never thought about firewalls or rate limiting or what it actually means to expose something to the internet. Omni forced me to learn all of it. I set up the entire infrastructure from scratch: Cloudflare, Nginx as reverse proxy, UFW as the firewall, Docker for the backend services, a separate PostgreSQL instance for the database. It also taught me real-world HLS video streaming and how to manage caching properly for a true end-to-end product. I learnt a lot. It's one of the projects I'm most proud of.

Built right after Omni, when I needed a way to monitor my VPS from the phone instead of opening a terminal every time. A Telegram bot with a menu-based dashboard to check VPS metrics, memory, Docker containers, and services in a couple of taps. It sends alerts when a container crashes, stops, or when a resource usage crosses a set threshold. Mute controls let me silence notifications during deploys or maintenance, so the bot stays useful instead of spammy.

Powered by Google GenAI API. It mimics human texting with natural message splitting, dynamic typing delays, and fully customizable behavior.

Supports image, audio, video, and document conversions with advanced options such as resizing, rotating, filtering, and quality adjustments.

Turn your Instagram DMs into an autonomous AI persona. Mimics human behavior with realistic delays, typing simulation, and a custom deepfake persona.

A custom website I designed and developed for a music group.

A fun plugin for Paper that spawns "Markolino" a bot that wants to kill you!

Keybind to switch texture packs on/off ingame.

Automatic content extraction. Supports batch scraping, extracts articles with full metadata, media and structure. Structured JSON output.

A fun snake game made with React!

A fun clicker game made with React!

A minesweeper game with cheats, because why not?