AI-Based Rubric Assessment Application

#

In this assignment, I built a small AI-driven application that gives a guiding assessment of a student report. The purpose of the project was not to create an automatic grading system, but to explore how a large language model can be integrated into a real software flow.

The application takes a student assignment as text input and returns structured feedback based on a rubric. The rubric is based on three source documents: the learning goals for the internship, the report requirements, and the Dare-Share-Care method. These documents are included in the prompt sent to the language model, so the generated feedback is grounded in the actual assessment material.

My Technical Choices

#

I chose to build the backend in Node.js using the built-in HTTP module. This kept the project simple and made the data flow easy to understand: the client sends assignment text to the backend, the backend builds a prompt, sends it to an external LLM API, receives the response, and returns structured JSON to the client. ß The application has three main endpoints:

• GET /rubric returns the rubric used for assessment
• GET /source-materials returns the markdown source documents
• POST /assess receives an assignment text and returns structured feedback

The LLM response is designed as JSON instead of plain text. This makes it easier for the frontend to display the result and makes the feedback more useful. The response includes an overall assessment, feedback for each criterion, strengths, weaknesses, improvement suggestions, questions for further dialogue, and uncertainties.

Prompt Design

#

I designed a system prompt that tells the model to act as a fair and critical teacher. It also makes it clear that the assessment must be guiding and not presented as a final grade.

The user prompt includes:

• the rubric
• the student assignment
• the source materials
• a required JSON structure

This helps reduce useless feedback and encourages the model to stay close to the learning goals, report requirements, and Dare-Share-Care method.

Reflection

#

One strength of this solution is that the assessment is transparent. Each criterion is connected to the rubric and the source material, which makes the feedback easier to understand and discuss. The structured output also makes the application easier to extend later.

A limitation is that the quality of the feedback still depends on the language model. The model can misunderstand the assignment, overlook important details, or give feedback that sounds convincing without being fully accurate. Because of this, the result should only be used as a starting point for reflection or dialogue, not as an automatic final assessment.

Project: https://github.com/MStensk/AI-LLM