Hi, I'm Lorenzo Giovanni Vecchio

A Rust clone of PocketBase, built as a learning project. Implements the full CRUD API for collections and records with a typed field schema, SQLite-backed storage (bundled, single binary), and a validation layer mirroring PocketBase's forms architecture. Planned extensions include a Rhai scripting engine for hooks, a web UI for managing collections, a CLI for pushing schema changes to a remote server, and a built-in MCP server for AI control.

A fully hand-rolled multilayer perceptron written in Go, with zero ML dependencies. Implements every component from scratch: forward and backward passes, ReLU and Softmax activations, cross-entropy loss, He weight initialisation, and mini-batch SGD. Trained on 60,000 MNIST images, it reaches 97.4% accuracy on the 10,000-image test set. The source code is intentionally written as a learning artifact — every concept (backprop, the chain rule, overfitting, normalisation) is explained in plain English directly in comments, making the codebase a readable reference for programmers new to machine learning.

A small desktop app I built at the request of some friends — with heavy AI assistance. It handles FatturaPA electronic invoices (XML, ZIP, or signed .p7m files): upload a batch, filter by supplier, customer, date, amount, and document type, then export the selection as a ZIP archive organized by supplier or customer. Sessions can be saved as projects and reopened later. It also features an AI analysis agent (Claude/GPT) that reads the loaded invoices and generates structured reports — with follow-up conversation support and Word export. Runs locally via Tauri: no data leaves your device unless you configure an AI API key.