1. The Mission

Most retro tools help you hunt for a specific thing: "Find me a Voodoo card, a 486 box, this exact sound card."

My real problem was different. I do not always want to buy something, but I always want to know:

• what showed up today
• what is flooding the market
• what is quietly disappearing
• where prices are drifting over time

Doing that manually means multiple tabs, multiple marketplaces, duplicated listings, and a vague feeling of "I think this is getting more expensive."

retro.dog turns that vague feeling into a quick daily radar: one place where you can see, in seconds, whether anything interesting is happening on the retro market today.

2. How it works (human language, not marketing copy)

Under the hood, retro.dog is a simple idea with a lot of plumbing:

1. A crawler (for now focused on eBay DE) collects raw listings: titles, descriptions, prices, locations, and images.
2. A small Qwen 4B model filters out non-retro noise so modern hardware and junk do not pollute the feed.
3. Mistral Nemo translates and normalizes messy seller text into something a model, and a human, can actually use.
4. A Gemma VL 12B multimodal model looks at images plus text and produces an assessment: condition, interest level, and how special something might be for collectors.
5. For some top picks, an extra layer turns these judgments into more readable summaries.

What you see on the site, the daily top pick, the curated feed, and the AI insights on a product page, is the visible part of this chain.

3. Planning on the go - when reality beats the spec

retro.dog did not emerge from a perfect spec executed by perfect agents.

I started with detailed PM documents and let AI agents generate a lot of code and prompts. On paper, it looked great. In reality, reality won:

• AI wrote beautiful, structured prompts for other models that rarely worked as well as they looked.
• Some factors that sounded clever in theory were useless on real listings.
• Early model choices and clean architectures broke as soon as they hit noisy marketplace data.

Most real progress happened in planning on the go:

• adjusting the plan whenever real data punched a hole in it
• throwing away or simplifying factors that did not survive contact with reality
• taking AI generated scaffolding, then hand-tinkering the parts that actually mattered

retro.dog is the opposite of a "we followed the spec" project. It is a feedback loop between idea, implementation, ugly data, and revised idea, with AI as a helper, not an architect.

4. What retro.dog is not (yet)

This is a public ALPHA, not a finished product. Important caveats:

• One source only (for now): retro.dog currently focuses on eBay DE. One stable source is better than four half-broken ones when you are solo and paying for proxies yourself.
• No deal score or price verdicts: the system does not tell you whether a listing is a great deal or a rip-off. Without deep categorization, cohorts, and proper time series, price intelligence is more theater than truth.
• Bare-bones search: search and filters are intentionally simple. No fancy search engine yet. The focus is curated discovery, not perfect retrieval.
• AI is helpful, not omniscient: Qwen sometimes filters out legit retro items, Gemma sometimes over- or underestimates interest, and some summaries are better than others.

I prefer an honest radar with clear limitations over a fake sense of precision.

5. Who is this for?

In this phase, retro.dog is mainly for:

• retro collectors who prefer a daily market pulse over endless manual hunting
• tech people and developers who enjoy seeing scraping, small models, and heuristics become a working tool
• AI tinkerers who are tired of polished demos and want to see what real product building looks like with imperfect models and constraints

6. Where this is going

The long-term direction is simple: from curated snapshot toward a true retro market pulse.

In practice, that means:

• adding more sources once economics and stability make sense
• building time series views and category trends instead of only daily lists
• experimenting with honest, data-backed price intelligence only when categorization, cohorts, and history are strong enough

Even as it grows, retro.dog will likely remain what it started as: a selfish, experimental tool built to understand a niche I care about, and a live testbed for how far small models, scrapers, and heuristics can go in the real world.