heyprof: An AI-Assisted Platform for Higher Education

Authors: Volker Reichenberger, Prof. Dr. Dirk Schieborn · NXT Sustainability and Technology, Reutlingen University · March 2026, Version 1.0

Ongoing project — development and evaluation are not yet complete. heyprof is in alpha stage. The features, evaluation results and conclusions described here reflect the state as of March 2026 and are continuously updated.

heyprof is a web-based platform for teaching organisation and course support, distinguished by the consistent integration of Large Language Models (LLMs). At its core is a dialogue-based AI tutor that guides students through Socratic dialogue — rather than providing ready-made solutions. It takes into account the context and current state of the course to provide targeted, level-appropriate support.

The Problem: AI in Teaching Without Context

Large language models can be used beneficially in teaching — but learners are often left alone with general AI assistants. Beyond the problem of cognitive bypass, there is frequently the issue that AI-generated solutions do not match students’ current level of knowledge. The AI does not know the current state of the course and therefore often delivers answers that are too complex or that hinder the learning process by using concepts not yet introduced.

heyprof addresses these requirements by having the AI support students in a context-sensitive and targeted way. The AI tutor knows the exact text of the current assignment, the model solution and the instructor’s private tutor notes, the student’s existing Python code, and relevant passages from the course script.

Technical Architecture

Frontend: Next.js 15 with React and TypeScript (App Router). Tailwind CSS and Radix UI primitives. Mathematics rendered client-side via KaTeX and MathJax.

Backend: Next.js API Routes (serverless) for all AI and business logic. Supabase (PostgreSQL with Row-Level Security) for data persistence and authentication.

AI Services: OpenAI gpt-4o-mini for chat, grading, translation and summarisation; text-embedding-3-small for semantic document search (RAG); Whisper v2 for lecture transcription; DALL-E 3 for automatic assignment cover images.

Storage: Supabase Storage for PDFs, images and audio files. Semantic embeddings stored in PostgreSQL’s pgvector extension. The application runs on Vercel and requires no dedicated server infrastructure.

Features for Instructors

Course and session management: Instructors create courses in the dashboard and receive an eight-digit access code. Each course is divided into sessions corresponding to individual lecture dates. The order of all elements can be adjusted via drag-and-drop.

Material management and annotation: Uploaded PDF slides can be annotated with a freehand drawing tool. Students see these annotations as an overlay. After upload, materials are ingested: text is extracted, split into sections, embedded and stored in the vector store, so the AI tutor can retrieve relevant passages from the course script on demand.

Assignment editor: The editor supports three types: free text, multiple choice and code (Python). For each assignment, instructors can add a model solution, private tutor instructions, labels and collections. Assignments can be imported from LaTeX source files. Cover images are automatically generated via DALL-E 3.

Exam management: Exams consist of numbered questions with point values and a countdown timer. Questions can be imported from scanned PDF exam sheets — an AI pipeline automatically extracts text, numbering and points. After the exam, all submissions are available with AI-generated assessments.

Recordings and transcription: A browser-based audio recorder enables direct recording of lectures. Whisper v2 transcribes the recording in the background. GPT-4o then generates a structured summary — visible to all students.

Features for Students

Dashboard and Learning Studio: The dashboard shows enrolled courses with progress indicators. The Learning Studio lists all sessions chronologically — with PDF viewer, annotation overlay, lecture recording with transcript and summary, and the session’s assignments on a single page. In the PDF viewer, students can select a region; the AI tutor receives this excerpt as a focus hint.

Assignments — six submission channels:

• Text/Code — direct input with immediate AI grading
• Drawing canvas — freehand solution for mathematical or diagrammatic answers
• QR photo — smartphone scans QR code, takes photo, answer appears on desktop
• AI tutor — immediate Socratic support in the integrated chat
• Practice variant — AI generates a structurally identical assignment with different numbers or function names
• Show solution — after a genuine solution attempt

AI Features and Pedagogical Integration

Retrieval-Augmented Generation (RAG): Before the AI tutor responds, its system prompt is enriched with the full assignment text, private tutor instructions, the student’s current Python code, semantically retrieved passages from course materials, and the instructor’s handwritten annotations. When a student asks “What is a linked list?”, the system retrieves the most relevant sections from the course script — not from the model’s general world knowledge.

Socratic mode and explanation mode: In Socratic mode, the tutor asks only counter-questions that move the student one step forward without revealing the solution. In explanation mode, the tutor explains directly, actively and vividly — with concrete examples and everyday metaphors, referencing the provided slide content.

Automatic grading: Text answers are graded via a structured GPT-4o call. Handwritten or photographed solutions are analysed by the vision model, checked for relevance, and converted into LaTeX notation.

Pedagogical Principles

• Constructive alignment: Every assignment, tutor instruction and collection is linked to a concrete learning objective. The AI guides students towards the expected solution — without giving it away.
• Zone of proximal development: The Socratic tutor always tries to identify the boundary of the student’s current understanding and ask a question that goes exactly one step beyond it.
• Deliberate practice: The variant function enables immediate practice after understanding the original — same concept, new numbers, immediate grading.
• Multimodal input: heyprof accepts typed, drawn and photographed input equally, reducing the gap between thinking and submission format.

Outlook

heyprof demonstrates that AI support in higher education does not have to mean students bypass the learning process. By embedding the AI tutor deeply into the assignment workflow, injecting pedagogical context, and restricting it to Socratic guidance, a potentially harmful tool becomes an effective learning aid.

The first courses using heyprof will take place in the 2026 summer semester. heyprof is already available as an alpha version at heyprof.app.