Agentic AI: Intelligent Study Assistant

Thesis Title: Agentic AI: Conception and Prototypical Implementation of an Intelligent Study Assistant Based on a Large Language Model

Original Title (German): Agentic AI: Konzeption und prototypische Umsetzung eines intelligenten Studienassistenten auf Basis eines Large Language Models

This repository contains the conception and functional prototype of an intelligent study assistant, developed as part of a Master's thesis at Hochschule Düsseldorf (HSD). The system uses a local multi-agent architecture to autonomously support student workflows such as deadline management and course organization.

🌟 Key Features

• Privacy-by-Design: The system operates entirely locally on consumer hardware (e.g., RTX 4090), ensuring sensitive data such as grades and appointments never leave your own infrastructure.

• Multi-Agent Orchestration: A central Routing Agent decomposes complex tasks and delegates them to specialized sub-agents for Moodle and calendar services.

• Standardized Interoperability: Uses the Model Context Protocol (MCP) for tool integration and the Agent2Agent (A2A) protocol for communication between autonomous agents.

• High-Performance Inference: Utilizes the vLLM engine with PagedAttention for highly efficient memory management and parallel request processing.

---

🏗 System Architecture

The system follows a hierarchical microservice structure that strictly separates cognitive planning from technical execution.

• Routing Agent: Acts as an orchestrator and uses the ReAct paradigm (Reasoning + Acting) to solve tasks step by step.

• Tool Agents (Moodle/Calendar): Encapsulate domain-specific logic and perform independent "sub-reasoning" (e.g., resolving course names to IDs).

• MCP Servers: Provide a standardized abstraction layer for technical access to external APIs such as Moodle or Google Calendar.

---

🖥 User Interface

---

📋 Prerequisites

Before setting up the project, ensure your environment meets the following requirements:

• Hardware: An NVIDIA GPU with sufficient VRAM (e.g., RTX 4090) is required to support the local inference of the Qwen3-4B model via the vLLM engine.

• Containerization: Docker and Docker Compose are necessary to orchestrate the distributed microservice architecture.

• Node.js: Version 22+ is used for the backend logic and the Next.js frontend.

• pnpm: The project uses pnpm for efficient package management within the Nx monorepo.

• Python: Required for executing the semantic evaluation pipeline and metrics analysis using the DeepEval framework.

• Moodle Access: A valid Webservice Token is required for course and assignment retrieval.

• Google Calendar Access: Configured OAuth2 credentials are required for managing calendar events.

---

🛠 Technology Stack

• Frontend: Next.js with real-time visualization of the reasoning process via Server-Sent Events (SSE).

• Backend: Node.js & TypeScript within an Nx Monorepo.

• AI Model: Qwen3-4B (Small Language Model), optimized for local execution.

• Validation: Strict runtime validation of all data structures using Zod.

---

🚀 Installation & Setup

1. Preparation

Install dependencies and copy the example configuration file:

pnpm i
cp .env.example .env

Enter your API keys and tokens in the .env file.

2. Network Configuration (Local Development)

To access the locally hosted Moodle instance, an entry must be added to the hosts file:

• Entry: 127.0.0.1 moodle
• Moodle will then be accessible via http://moodle.

3. Start System

Starts the entire agent system along with a local Moodle instance:

bash scripts/start-docker.sh

4. End-to-End Tests

Starts the system including WireMock to simulate the Moodle and Google Calendar APIs:

bash scripts/start-docker.sh test

---

🧪 Testing

Run tests for individual components:

# End-to-End tests
pnpm test:e2e

# Agent Framework unit tests
pnpm test:agent-framework

# Moodle Agent unit tests
pnpm test:moodle-agent

# Moodle Agent E2E tests
pnpm test:moodle-agent:e2e

---

📊 Evaluation & Research Findings

A quantitative ablation study examined the impact of explicit planning modules on Small Language Models (SLMs).

• Results: An explicit TODO list led to a performance drop of 10.7 percentage points compared to the baseline (56.0% vs. 45.3% pass rate).

• Inverse Scaling: The more complex the context becomes through planning patterns, the more the precision of smaller models decreases.

• Conclusion: For local agent systems based on SLMs, a "Less is More" strategy in prompt engineering is more effective than complex planning patterns from larger models.

Running Evaluation Manually:

pnpm run test:evaluation
# Start analysis script
cd evaluation
source venv/bin/activate && pip install -r requirements.txt && python evaluate_experiments.py

---

🔍 Transparent Open Glass Approach

The frontend provides a transparent "open glass" view into the agent's reasoning process. Users can observe in real-time how the system decomposes tasks, delegates to sub-agents, and executes tool calls. This makes the decision-making process fully transparent and comprehensible.

---

👨‍💻 Author

Malte Sehmer, Hochschule Düsseldorf

Submission Date: December 23, 2025