retro.dog – AI-assisted market radar for retro computing

retro.dog is an independent public project: an AI-assisted market radar for retro computing listings.

It monitors marketplace listings, filters obvious noise, enriches inconsistent product data, and turns raw listings into signals that collectors can review.

During its first public alpha test, the system collected and processed around 15,000 retro computing listings.

Marketplace data is messy by default.

Titles are inconsistent, categories contain a lot of noise, descriptions may be incomplete or multilingual, and images do not always reveal the details that matter. Scraping can fail, while proxy and bandwidth costs grow quickly at scale.

AI helps with parts of this work, but it also misses subtle items, assigns confident labels to poor results, and cannot reliably decide what collectors will find interesting.

The pipeline therefore cannot simply ask a model to choose the best listings. It needs several filtering stages, cost controls, review steps, and human judgment.

I built a pipeline that crawls listings, extracts product data, filters obvious noise, normalizes inconsistent text, runs AI-assisted assessments, generates summaries, and presents selected items through a public interface.

Where possible, cheaper filtering happens before more expensive processing. Smaller models are used for tasks where they perform well enough.

Human review remains between the model output and anything that is published.

This is practical AI work under real constraints: noisy data, scraping failures, model limitations, cost, review, and public-facing output.

The goal is not to let AI decide everything. It is to build a workflow where AI adds useful structure while human judgment remains responsible for the final result.

Laravel, Livewire, Puppeteer, MariaDB, scraping, marketplace data, LLM pipeline, filtering, normalization, human review, cost-aware AI workflow.