1,635 conversations triaged, 174 filed for a human reply, zero posted by a machine.

Radar is an open-source Kubernetes UI with a small founding team and a community that talks about it in eight places at once: Reddit, Hacker News, X, LinkedIn, dev.to, YouTube, GitHub, and the long tail of blogs. The founders were finding mentions days late or not at all, category questions ("Lens alternative?", "what do you use instead of the Kubernetes dashboard?") went unanswered, and a manual sweep cost an hour a day nobody had. The one thing the team did not want was a bot posting on their behalf.

A daily listening system, in production from 7 July 2026. Eight channels are swept every morning with per-channel keyword sets and a deduplication store, so an item reaches the team exactly once. Every hit is classified as a brand mention, an own post, or a category conversation; own posts are excluded, brand mentions get a 21-day window and go to the top. An LLM triages each item against a verified facts pack and the founders' reply guidelines, returning relevance, priority, a strategic angle, and a first-draft reply in the founder's voice. Engage-worthy items are filed into Linear with an owner and priority; the rest are dropped with the count shown. The whole thing lands as one HTML email, with per-channel status so a silent failure is visible. Nothing is ever posted by the machine: every draft is edited and sent by a human from their own account. When four LinkedIn brand mentions were found to have slipped through in August, the detection layer was rebuilt and eleven missed posts were recovered. In parallel, the first AI-visibility baseline for the project was run and the pages feeding competitors' AI answers were fixed.

The product demos were built the same way, as a build artifact rather than a recording. Six clips of 20 to 40 seconds (install, topology, a Helm rollback, an admission-rejection dead end, an agent diagnosis, and a one-minute timeline piece for the launch post) render from a single make target against a deterministic stage: a kind cluster with twelve real microservices, three deliberate faults to give a visibility tool something to see, a genuine two-revision Helm release to roll back, and eighteen KWOK nodes across three zones so topology and cost views look like a fleet instead of a laptop. Terminal takes are VHS tapes, UI takes are Playwright with a synthetic cursor, and each capture writes a cue file that the compositor uses to place captions, zoom and narration, so nobody opens a video editor. Every clip's script lives in version control next to a claim-check list that is verified against the current release before anything publishes. When the product changes, nothing is re-recorded; it is re-run. Running the render in CI on release tags is written but has not executed yet, so I do not claim it works.