Projects

Desktop comic & manga reader

The problem

Desktop CBR/CBZ readers are either heavy, forget where you left off, or burn your eyes during long sessions.

What I built

A Rust core handles archive extraction and paging without blocking the UI, with the library persisted in SQLite and organised into pockets. The interface is React on Tauri 2, with themes built for reading at night — sepia, dark and OLED.

Where it landed

A cross-platform desktop app that resumes reading where you stopped and keeps the library organised without relying on any external service.

Elegant résumé builder

The problem

Keeping a CV in LaTeX gives a flawless result but locks out anyone unwilling to fight the syntax.

What I built

I turned my own LaTeX template into a web app: sections you add and reorder on the fly, an A4 preview faithful to the printed result, and PDF export in one click. Everything runs client-side, with no sign-up and no server.

Where it landed

A public tool anyone can use, with the same typographic result as the original.

Anki add-on for vocabulary

The problem

Capturing vocabulary on the go is easy; translating every card by hand afterwards is what makes people abandon the deck.

What I built

A Python add-on that walks every note in a deck and fills the translation field in one pass, against a free translation endpoint with no API key. Language pair and field mapping are configurable.

Where it landed

A deck of hundreds of cards gets translated in a single click, with no credentials and no paid service.

End-to-end machine learning pipeline

The problem

Picking an algorithm out of habit is the fastest way to get a model that looks good on paper and fails in production.

What I built

A full pipeline over one operational dataset: preparation, feature engineering and a comparison of six algorithms — logistic regression, decision trees, KNN and others — all evaluated under the same protocol.

Where it landed

A reproducible comparison showing what each family of models buys you and at what cost, instead of a single model with nothing to contrast it against.

Document classification from raw PDFs

The problem

A long regulation in PDF is unstructured text: finding which article applies to a given case means reading the whole thing.

What I built

A system that ingests the PDFs, extracts and normalises the text — cleaning, tokenisation, stopword removal — and classifies each section into predefined categories with scikit-learn.

Where it landed

The document goes from a wall of text to a structure you can query by category.

Polyglot microservices architecture

The problem

Real distributed systems are rarely written in a single language, and microservice theory gets interesting exactly when different stacks have to talk to each other.

What I built

Three independent services, each on the stack that suits it — authentication in Go with Gin and Redis for sessions and tokens — coordinated by a central API Gateway with automatic service discovery.

Where it landed

A working testbed where service discovery, gateway routing and specialised databases coexist in one system.