arrow_backBack to Projects
rbAI: Senior Yr. Thesis icon

rbAI: Senior Yr. Thesis

AI & AgentsWeb Apps
Jun 25, 2026visibility3

An educational coding environment that monitors engagement behavior with a guided AI assistant for academic use.

ReactTypeScriptFastAPIPythonLangGraphDockerPostgreSQL

The Problem Statement

Abstract (taken from my thesis):

The use of Generative Artificial Intelligence tools in education is becoming increasingly apparent each year due to the benefits they offer. It can provide fast, easy access to information; however, such accessibility may hinder the learning process and introduce risks, such as dependency, reduced critical thinking, and concerns about academic integrity. Such risks are evident with programmers who are still learning fundamental skills. Their early contact with AI tools led to reliance on generated output rather than encouraging independent problem-solving, weakening long-term understanding. With AI capable of producing near-accurate outputs with such ease, it is becoming hard for instructors to know whether such work indeed reflects genuine learning. As a result, assessment methods based on outputs are becoming insufficient, showing the need for process-oriented assessment and a new definition of AI’s role in learning. This study addressed these challenges through rbAI, a coding playground integrated with engagement monitoring and a guided AI assistant. It introduced an engagement model that monitors key interaction patterns to evaluate the student engagement while reframing AI as a thinking partner that guides learners through problem-solving without promoting cognitive shortcuts.

Motivation of Creating this App

The use of AI now in education do indeed provides significant benefits, however, this same access hinder the learning process. The problem in such tool is that, it makes doing things "easy." The hard part of actually doing something somehow being done now by AI. You think that would be a good thing, but in reality, it is not. Because learning is not easy to begin with; it is a product of trial and error that we experience when we are learning something. It is in this very struggle that we actually nurture our minds and enhance our critical thinking.

Think about it, if a student uses AI on their assignments or learning tasks, and got a high remarks out of it, who is the teacher actually giving grades to? The AI or the student? And if the former, what the fuck happened to learning? I mourn for the future humanity

This is the problem I want to solve.

I want to protect the integrity of learning.

That is why I created rbAI. To somehow protect the learning of students, and specifically, targeting lower years so that they can build enough foundation and most importantly, so that they can audit the output of a GenAI.

Solution

Since i'm trying to give a solution of that problem mentioned above in my field, the closest thing that mostly encapsulates Computer Science is coding or programming. With that in mind, I chose a coding playground like leetcode, codewars, etc..

I designed the solution with two main features:

  • A process-oriented design so that I can look at the process of the output of the student to determine whether they do it on their own or they have external help (i.e, AI)
  • A harnessed AI agent that acts as guide for the students without spoonfeeding them.

Architecture & Tech Stack

To handle real-time telemetry processing alongside autonomous AI planning and safe code execution, the system is decoupled into three core services orchestrated via Docker Compose:

  1. Frontend Client: Built with React, TypeScript, and Vite. It utilizes a Monaco Editor for the coding playground and captures high-resolution telemetry events (focus changes, keypress intervals, run attempts).
  2. Backend API & Data Engine: A FastAPI application in Python backed by SQLite (for prototype) and PostgreSQL (in production). This handles session coordination, data fusion for the behavior engine, and hosts the LangGraph chat runtime.
  3. Execution Service: A completely isolated microservice (using docker) specifically for running student code submissions safely against validation test cases.

The Challenge

The most complex technical hurdle was bridging the gap between raw UI telemetry and meaningful pedagogical intervention. I needed to quantify "learning" versus "slacking" or "cheating."

1. The Behavior-Driven Engagement Model

To translate behavior into actionable data, I built a data fusion engine that calculates an Engagement Score. This pipeline filters out noise (like spamming random keys to inflate metrics) and distinguishes between actual off-task behavior and productive, reflective pausing.

2. Socratic Pedagogy

Even with perfect behavioral insights, standard LLMs tend to be overly helpful. With that, I implemented a strict Pedagogical Firewall inside the LangGraph runtime. Before any prompt reaches the core AI, the system analyzes the student's request to select an appropriate response. This allows the system to refuse to provide direct answers or code solutions, and only allow guided response to the students to encourage them to solve the problem themselves without spoonfeeding them the solution.

Through this workflow, the system's AI assistant guides the student with their task at hand augmenting their human capabilities. In this way, the system is not promoting overreliance and total replacement to the students cognitive load.

Lessons Learned

Building rbAI taught me the intricate realities of applying AI to human-centric problems. I learned how to translate highly subjective states (learning, struggling, reflecting) into resilient, deterministic algorithmic pipelines that can't be easily gamed. Furthermore, integrating LangGraph completely shifted my perspective on modern LLM development (this was my first time creating a system involving orchestrating such complex workflow!). Moving away from standard conversational endpoints, I learned how to build true autonomous agents with explicit state tracking, bounded tool execution, and complex routing architectures. Ultimately, I proved to myself that sometimes the best way to leverage AI isn't to build a tool that does the work for you, but one that effectively challenges you to do it yourself.

Note: This software is only intended for students in programming education to somehow harness the ongoing risk of using AI in education. Since students, such as novices, are the ones who are most affected, unguided use of AI replaces their productive struggle. This system aims to mitigates the risk of overreliance and dependency to the AI through process-oriented evaluation in learning task instead of product-oriented and responsible use of AI fit for academic use.

Quokka AIQuokka