Bollette Calcio

The problem

A personal project started in January 2021: a web app for betting on football matches with virtual money, with a ranking of users, and later slot machines and multiplayer poker. It stayed alive for about five years and is split across six repositories, so it is mostly a story of how the architecture changed.

What I built

• v1 (2021-2022): an Express, MySQL and Redis API with a React UI. Bets were settled by a Python scraper spawned as a child process, with tokens stored in Redis.

• v2 (2023 onwards): the scraper became a separate Flask microservice and the backend was rewritten in NestJS with TypeORM, JWT and Swagger. Poker runs over WebSockets, with the table modelled as an XState state machine, bots and an event log.

• Odds model: a neural network with team embeddings that generates Serie A odds, so the app does not depend on a paid odds API.

• Agent (2025 onwards): an assistant that fills the betting slip from sentences like "checkout 20". The UI moved to Vite, React 19 and Tailwind 4.

Interesting decisions

The agent went through two designs. The first, built with LangChain and OpenAI, was a pipeline of specialised steps: intent detection, team extraction with a structured parser, fuzzy match of the fixture with embeddings and a similarity threshold, then outcome and amount. The second, from October 2026, replaces prompts and embeddings with typed questions (choices and yes/no) asked in parallel to a decision model. It returns probabilities rather than text, so replies are templates. It uses confidence thresholds and falls back to the first extractor when unsure.

The betting slip is shown in the UI and the checkout stays an explicit action by the user, so the human confirms before money moves, even virtual.