Inspiration
Editing a short-form reel from raw footage takes hours of manual cutting, captioning and voiceover. We asked: what if a Software Factory could build and continuously improve the video editor itself, and prove every change by rendering a real clip?
What it does
FrameFoundry turns raw footage + a goal (e.g. "30s Sedona travel teaser") into a finished, judged vertical (9:16) clip: shot selection, cuts, text overlays, ElevenLabs voiceover, music and captions, then an automated QA/judge pass that scores the result.
How we built it
- Built by the 8090 Software Factory: planner, coder, reviewer and eval agents ship PRs (work orders) to a code-driven video editor. No product code was hand-written.
- Every PR is judged by rendering a real test clip: the eval agent renders, scores (QA + LLM judge), and only better clips ship.
- Hybrid deterministic + LLM: ffmpeg does the deterministic rendering; OpenAI models plan shots and judge; ElevenLabs generates the voice. Costs stay minimal.
- Scoreboard: cost per shipped clip, QA/judge scores, and run history, live at https://framefoundry-fawn.vercel.app (/studio, /pitch).
Challenges we ran into
Legible text overlays on vertical video, keeping per-clip LLM cost low, and making the judge reliable enough to gate merges.
Accomplishments that we're proud of
An end-to-end factory loop: work order -> PR -> rendered clip -> judged score -> deploy, running overnight on a Mac mini.
What's next
Full-length 16:9 4K vlog editing from raw footage, and more agent skills for production-grade reels.
Team
Rahat Khanna (solo human), with Grok Bot as AI teammate (steered the 8090 Software Factory build).
Code (MIT): https://github.com/mappmechanic/frame-foundry