What they did
Daniel, who says he is bad at naming things, built namematch for a dev.to open-source AI challenge. You paste a GitHub repo link or a markdown file and answer a few questions. A supervisor agent reads the project, fills in the questions and chooses which of three namer agents to run (professional, short, playful). A domain-checker agent suggests variants for taken names, while plain code asks the registries over RDAP whether .com, .dev, .app, .tech or .ai is free. A feedback agent then reworks the list from chat. The six agents are built with Mastra and run on Gemma 4 through Google's free tier, with fixed workflows deciding the order of steps rather than letting agents call each other. Every call is traced in Sentry.
How it went
Early runs failed on the free tier's rate limit until retries waited as long as Google asked. Traces showed the namers spent most of their time thinking: switching thinking off cut one namer call from 79 seconds to 5, and a full naming round dropped from 8 to 10 minutes to 2 to 5. He then pointed the app at its own repo with only .tech allowed. After seven rounds of steering, it offered NameMatch with a free .tech domain, and the app was renamed from Nammit to namematch.
Worth knowing
Thinking stayed on for the supervisor and feedback agents because names got worse without it, so those steps remain the slow part. The hosted version sends your project description to Google's API; running Gemma locally through Ollama is possible but he had not tested it.