Software project
barato.si
A daily index of US price drops and probable price errors: a static Nuxt site fed by a TypeScript pipeline that keeps its own price history.
My contributionProduct design, data contract, and pipeline architecture
Built withnuxt / typescript / data-pipeline / statistics / forecasting / static-site
About the project
barato.si publishes the deepest US discounts of the day and the prices that look like mistakes. It is written in Spanish, with an English edition at /en. The site is static Nuxt on GitHub Pages, and a TypeScript pipeline collects the data and rebuilds it every three hours. It is a Formship LLC project.
Where the data comes from
The pipeline reads first-party catalogs from stores like Steam, GOG and Epic. It also reads deal aggregators under their terms: DealNews, Techbargains, CheapShark and CamelCamelCamel. It doesn’t scrape retailer pages and it identifies itself honestly. When a source blocks automated requests, that source stays out.
Judging a price
Every collected offer goes into a price ledger keyed by product. The ledger keeps each product’s daily low for 120 days, including offers that never get published. A price is judged against the strongest reference available: the product’s own observed median first, then the same model at another store, then a dated curated price, and the seller’s list price last. A list price alone can’t flag a probable error, because sellers inflate them. Every point of the error score is stored with the reason behind it.
An experimental forecast
The newest piece asks one question per product: will a better price show up within 14 days? It fits an empirical-Bayes beta-binomial model to the ledger, so a product with little history leans on the rest of the data. The joke is that .si stands for super intelligence. The model has to earn the name in public: its Brier score is published next to a naive baseline, along with calibration buckets that compare what it predicted with what happened.
The habit of scoring a predictor against a naive baseline, and of showing people how far to trust it, comes from my master’s coursework in applied machine learning and human-centered AI.
What it doesn’t claim
Feeds are downstream of whoever posted the deal first, so the index is never first. Merchants often cancel price errors. Reference prices are evidence, not verified market values, and the page tells readers to confirm the price at the store. An empty error list is a legitimate result, and the thresholds don’t drop to fill it.