AI manager talks a person through their grading verdict, out loud, pointing at the file

A voice-based feedback flow that runs right after a piece of submitted work gets graded (see the automatic grading project). The same AI manager the person has been working with picks up that grading result and walks them through it conversationally: what the issues are, what was missing, what they did well, and how to go beyond the bar. Each point ties back to a specific highlighted region on the file they actually submitted, so the person can see exactly what the manager is talking about while they listen.

A written grading report is a wall of text. Most people skim it, miss the actual point, and stay stuck on the same issues next time. Having the AI manager talk them through it, out loud, pointing at the actual part of their work that needs the fix, is how someone actually learns from a verdict instead of bouncing off it. Keeping it the same AI manager the person has been talking to the whole time, instead of a separate feedback bot, makes it feel like part of an ongoing conversation, not a report thrown over the wall.

The walkthrough consumes the structured verdict directly from the grading system, so it never re-grades the work or wastes effort regenerating a verdict that already exists. Each item in the verdict (issues found, style notes, what went well, how to exceed the bar) is routed through a prompt that puts the AI manager in the right voice for this scenario and writes a short spoken script per item. Each script is paired with a highlighted region pulled from the original submitted file (a PDF or image), so the audio and the visual highlight stay in sync as it plays. To keep initial cost down, only the issues are narrated up front. The positive points are stored and only narrated on request, generated the moment the person actually asks to hear them. The whole flow runs inside the same conversation system as the rest of the person's work, so the manager keeps the same voice and memory throughout.

Major contributor on the walkthrough prompt, its rewrite, the evidence-highlighting layer that ties each comment to a region on the original submission file, and the on-demand positives split that defers cost until the person actually wants to hear them. Audio generation itself is owned by another contributor.