Technical teams breaking out now in AI, robotics, defense, energy, semiconductors and manufacturing. Ranked by code, not by a model. Every sentence links to the line it came from.
Five steps, one command, once a day. The model never ranks anything and never gets the last word.
New GitHub repos in thesis topics, Show HN and Launch HN posts, arXiv papers that ship code, and the current HN hiring thread. Every record keeps its source URL and fetch time.
Records join into teams through shared keys: a GitHub org, its website, the repo a post or a paper links to. One team, many signals.
Plain arithmetic ranks the teams: star velocity, launch traction, recency, founder signals in bios, independent sources, signs of a company forming, thesis fit.
The model reads only the fetched sources and writes short claims, each with a word-for-word quote. Code checks every quote. A claim whose quote is not there is discarded and counted.
This page, the JSON behind it, and an MCP server. Off-thesis teams are listed, not hidden. Unverified claims never appear.
Counts from the run that built this page.
Scored high enough to be read, then judged outside the thesis: consumer apps, developer tooling, research code with no team.
Scored high enough to be read, but nothing the model said about what they build survived the evidence check.
Radar is an MCP server. Public, read only, no key.
In Claude, add the URL under Settings, Connectors, Add custom connector. Tools: latest_companies search_companies get_company run_stats