svenhq

From raw footage to posted video

The AI editing system behind Impact Avalanche’s client videos. Clips go in one end; edited, captioned, approved posts come out the other. A Mac at home does the work, Claude makes the editing calls, and a person approves every video before it posts.

    The stack

    Most of it runs on one Mac. The cloud pieces hold the files, the database, the review page and the posting. Claude is called only for judgment; everything repeatable is code.

    JobWhat does itWhere it runs
    Upload and reviewA small web app: an upload page, a review page with notes and approvals, a plain-English footage search, a posting calendar.Next.js on Vercel
    FootageEach client’s shared drive: an inbox, raw shoots, edits, a B-roll library.Google Drive
    The workerA MacBook Pro (M2 Max) that never sleeps. Runs the engine at 6 AM, 10 AM, 2 PM and 6 PM, and big jobs overnight.At home
    TranscriptionWhisper large-v3 turbo, word-level time stamps.On the Mac, offline
    JudgmentClaude, run by Claude Code without a screen: plans, edit recipes, captions and posts, revisions from notes, B-roll tagging.Called from the Mac
    Cutting and renderingPython and FFmpeg, encoded on Apple’s hardware video encoder.On the Mac
    ReframingApple Vision face detection keeps the speaker in a vertical frame, with a lazy camera so it doesn’t twitch.On the Mac
    SoundDeepFilterNet noise removal, hum removal, and a studio EQ picked by ear.On the Mac
    Quality checkA scripted checklist plus a blind Claude Sonnet judge, held to “flag only” until it agrees with a human 8 times in 10.On the Mac
    StateEvery shoot, clip, video, post and note, with a nightly backup.Neon Postgres
    PostingAyrshare schedules and posts to each client’s own accounts, then syncs links and views.Cloud
    Podcast episodesThe same engine, long form: a cold open, a tightened body, a mid-roll, an end card and chapters.On the Mac

    Where people stay in charge

    The machine does the hours. People make the calls that matter, at four fixed points.

    What to make
    Each short gets a style before it’s cut. A tagged upload or the client’s house style goes straight through; anything ambiguous waits for the team to pick.
    Every video
    Nothing posts until a person approves the video and its captions. Notes go back in as a new version.
    First posts
    A client’s first post on each platform waits for a yes. The rest follow once it has gone out cleanly.
    The switch
    Posting is off for a client until someone turns it on, and a test mode logs what it would have sent.

    What it never doesDelete or move footage outside its filing step. Overwrite a version. Post anywhere but the client’s own accounts.

    Why it’s built this way

    AI makes the calls; code does the work. Claude decides what to cut and why. Plain code does the cutting, the same way every time, so any render can be checked, repeated and fixed.

    The heavy lifting runs on a Mac at home, so hours of footage cost nothing per minute to transcribe or render. And it’s part of a bigger system: the same Claude that runs this engine runs inside SI-OS, the AI team behind the rest of the work.