Janet A. Carr’s first-ever sponsored stream (partnered with Render) is a 6:53:57 build-in-public marathon: she and “a friendly clanker” (Claude Code) build an AI auto-clipper that turns a VOD into topic-matched clips, and against all live-stream odds the thing works and gets deployed to Render before the stream ends. The description teases “45 minutes of clip farming and time wasting to give haters some ammo before actually getting into the coding.”

The plan

  • Upload a VOD to a FastAPI web service, then hand the rest to Render Workflows.
  • FFmpeg pulls the audio, splits it into ~10-minute chunks, and re-encodes at a low bitrate.
  • OpenAI Whisper transcribes each chunk with timestamps (fan-out), then the LLM glues the chunks together.
  • An LLM scores the transcript for clip-worthy, prompt-matched moments and returns timestamped candidates with a clip score.
  • Parallel FFmpeg jobs cut the real clips out of the original VOD; the whole pipeline is orchestrated in Python via the Render Workflows SDK and shipped as a render.yaml Blueprint.

The build

  • Sponsored-stream chaos: broken credit links, a Homebrew ownership meltdown, “S3 bucket not configured” → temp-disk streaming upload, OBS/macOS screen capture dying mid-stream.
  • Whisper fan-out of 10 parallel tasks hit rate limits → capped at 2 concurrent transcriptions.
  • OpenAI billing surprise: ChatGPT premium does not include API credits → added $20; transcribing 1.5 hours of audio ended up costing $0.59.
  • A 6GB VOD broke the web-service upload (413 / stall at 36%) → FFmpeg re-encode to 500MB sailed through.
  • Print debugging, Emacs, Ghostty window rage, and Siri activating every 30 minutes.

What was actually delivered

  • Working end-to-end pipeline: process VOD → probe VOD (FFmpeg audio extraction) → transcribe chunk → score clip window → render clip, visible task-by-task in Render’s workflow dashboard.
  • Topic-driven clipping instead of viral-moment hunting: a user prompt (default: “strong opinions, rants, jokes, and explanations about programming software development…”) plus a per-candidate clip score — 33 candidates found, top 5 selected.
  • Deployed live to Render via Blueprint (“VOD clippers live”), re-tested on the deployed instance with real clips rendering into the page.
  • Open-sourced repo with a Deploy-on-Render button; she suspends the service and rotates the API keys before signing off.

“It pulled five clips. It chopped them up perfectly.”