Evalutron

Context

A small teaching team needed to evaluate reasoning-heavy assignments for approximately 210 students. Manual review created a two-to-three-week grading backlog and delayed useful feedback.

Problem

Free-form AI grading would not be dependable or academically appropriate. The system needed instructor-defined rubrics, known valid approaches, structured explanations, secure access, and final judgment by teaching staff.

What I built

After discussing the idea with the professor and grading team on a Friday, I built the first working version in two days. The Next.js application used an OpenAI API integration with a GPT-4 agent, accepted ZIP uploads, generated structured Excel reports, ran on Vercel, and restricted access to university VPN/IP addresses. It used no permanent file storage.

How it works

1. 01Upload student submissions as a ZIP archive for processing without permanent storage.
2. 02Provide each problem's rubric, expected reasoning, and two to four valid solution variants to the GPT-4 evaluation agent.
3. 03Compare the student's reasoning and final answer with the controlled references and generate a proposed score with an explanation.
4. 04Place suggested scores, rationale, and cases needing closer attention into a structured Excel report.
5. 05Require a teaching assistant or grader to compare every proposed evaluation with the original submission and make the final academic decision.

Outcome

• Delivered a working first version in two days and moved it into real use within two weeks.
• Helped the team recover from a two-to-three-week grading backlog.
• Two of four assignments were returned near the intended deadline; remaining delays fell to approximately four to five days.
• Continued to be used by the Computer Science department after I left the teaching-assistant role.

What I learned

Controlled evaluation works best when prompts behave like programs: clear rubrics, bounded solution paths, structured output, explicit uncertainty handling, and mandatory human review.