CodeSage

Most interview prep tools grade whether an answer is technically correct. They don't explain why an answer felt weak in the room: rambling, no clear structure, or missing the trade off a real interviewer would push on next.

Looked at how real interviewers actually score answers: structure, clarity, trade off awareness, and communication under pressure, not just whether the algorithm was correct. The gap wasn't grading accuracy. It was giving feedback specific enough to actually change what someone does the next time.

Built an LLM based evaluator that scores a response against a structured rubric (clarity, structure, completeness, trade off reasoning) instead of a single pass or fail, and ties each piece of feedback to the specific part of the answer it's about.

Early feedback was too generic to be useful. 'Be more concise' or 'explain your reasoning' are notes anyone could guess without reading the actual answer. The fix was forcing the model to quote the specific line the feedback referred to, which made the output something a person could act on instead of a polite platitude.

A working prototype that returns per answer, rubric based feedback instead of a single generic score.

Feedback that can't point at something concrete doesn't help anyone improve. Generic encouragement and a real critique look similar on the surface. Only one of them changes behaviour.