AccessCast

How it works ยท one clip, step by step

From one video to five.

Every step below shows the real output for one clip. Nothing on this page was written by hand.

  1. Listen

    Whisper writes down every word. A second model pins each word to the moment it is said.

    • Find the pauses

      A description has to fit between the words, so the code measures every pause.

    • Look

      Gemma 4 31B, on AMD, looks at still frames from the scene and lists only what it can see. Each fact names its frames.

      • Write

        Gemma writes one line from that list, and nothing else, inside a word budget.

        Gemma's first draft

      • Check

        Code checks the line, rule by rule. The model never decides whether its own line is good enough.

      • Repair

        The line goes back to Gemma with the exact rule it broke. Then the code checks the new line again.

        1. Fit it in

          A line is spoken in a pause. Where there is no pause, the extended version holds the picture while it is read.

        2. A person decides

          Every description line waits for a person. They see the frames, the checks and every repair, then approve, edit or reject it. Nothing is published while a line is waiting.

        3. Every way to watch

          The same analysis makes every version, so they never disagree.