Name: Harshit Verma
Email: [email protected]
GitHub: therealharshit
Organization: Sugar Labs
Project: LLM-powered Debugger for Pippy
Mentors: Walter Bender, Ibiam Chihurumnaya, Kshitij Shah
Debugging is one of the most challenging parts of programming for beginners, especially children who are just starting their coding journey. While many Large Language Models (LLMs) are marketed as tools that “write code for you,” they also have great potential to guide learners through the process of understanding and fixing errors in their code.
The Pippy Debugger project was developed to harness this potential by integrating an LLM-powered debugger into the Pippy Activity which is a Python learning environment in Sugar. The goal was not just to provide solutions, but to teach children how to debug, giving them context, explanations, and guidance in an age-appropriate manner.
- To integrate an LLM-powered debugger within the Pippy Activity.
- To make debugging accessible, engaging, and educational for children.
- To provide contextual feedback that helps learners understand the purpose of their code before debugging.
- To ensure seamless integration with Sugar AI, allowing Pippy to use centralized inference endpoints.
- To design a system where debugging history is saved to the Journal, enabling reflection and learning over time.
Many debugging tools are built for professional developers and are often too complex for beginners. Existing LLM applications primarily focus on code generation, while their educational potential for debugging remains underexplored.
The proposed Pippy Debugger builds on the idea of guiding learners through a debugging journey rather than spoon-feeding solutions. By integrating with Pippy and Sugar AI:
- Learners see the context of their code before debugging starts. .
- The output is filtered and age-appropriate for children.
- Python – Core programming language.
- FastAPI – Backend framework for the
/debugendpoint. - LangChain – To chain multiple LLM calls for better workflows.
- Hugging Face Transformers – For integrating Mistral 7B.
- Google Gen AI – For running Gemma-3-27B-IT.
- GTK 3 – For UI integration in Pippy.
- Sugar AI – Central inference system connecting Pippy with LLMs.
- Custom Markdown Parser – To render responses in the Debug Terminal.
- Learner writes Python code in Pippy.
- On clicking Debug, the code is sent to the
/debugendpoint in Sugar AI. - The backend processes the request, retrieves model inference, and formats the response.
- The response is contextualized, simplified, and displayed in a debugging terminal inside Pippy.
- Debug history (code, errors, suggestions) is saved into the Journal for future reflection.
- Studied the Pippy codebase and its integration with Sugar.
- Explored Sugar AI and tested existing endpoints.
- Evaluated multiple LLM options (CodeLlama, Mistral via Ollama).
- Mentor suggested moving inference to Hugging Face models.
- Improved base prompt design for debugging tips.
- Experimented with different ways of displaying debugging responses.
- Set up a FastAPI server with a
/debugendpoint. - Integrated Mistral 7B model through Hugging Face for debugging.
- Added Run and Debug buttons in the Pippy UI.
- Built a debugging terminal in GTK.
- Implemented contextualization – showing learners the code context before debugging starts.
- Raised a PR in Sugar AI for
/debugendpoint. - Learned LangChain for chaining LLM calls into a structured workflow.
- Prepared and presented midterm PPT for DMP evaluation.
- Continued developing and refining the
/debugendpoint in Sugar AI. - Completed integration of
/debugendpoint. - Tested debugging responses on various buggy Python codes.
- Found that some responses were too technical → began refining prompt workflow.
- Shifted to a larger model (Gemma-3-27B-IT) for more age-appropriate responses.
- Improved prompt workflows to simplify explanations for children.
- Enhanced Markdown parser for better formatting in the debugging terminal.
- Began work on saving debug history to the Sugar Journal.
- Designed a JSON-based structure for storing session details (code, error, AI suggestions).
- Partially implemented prototype for saving and retrieving debug logs.
-
Updated Pippy UI:
-
Debugging terminal with LLM response:
-
- Sugar AI: Add /debug endpoint for Pippy Debugger #28
- Pippy: Pippy Debugger Integration #113
- Help Activity: Add Debugger docs to Pippy #121
- Gained hands-on experience with FastAPI by building and refining the
/debugendpoint. - Learned to integrate LangChain for chaining multiple LLM calls and improving prompt workflows.
- Worked with Hugging Face Transformers and Google Gen AI APIs to run and test LLMs like Mistral 7B and Gemma-3-27B-IT.
- Built a custom Markdown parser and explored GTK 3 widgets for UI integration in Pippy.
- Improved understanding of designing API request/response schemas to enable clean communication between Sugar-AI and Pippy.
- Implemented error handling, response validation, and threading for a more reliable and responsive system.
- Learned the importance of refining prompts to make debugging responses child-friendly and age-appropriate.
- Experimented with contextualization and code-first workflows to guide learners before debugging.
- Learned how to dockerize applications and deploy them on AWS EC2 for scalable and efficient model inference.
The Pippy Debugger successfully laid the foundation for the LLM-powered debugger tailored for children. With the ability to display contextual explanations, simplify errors, and guide learners through debugging, the project moves one step closer to making programming education more accessible and engaging.
This project was completed under the mentorship of Walter Bender, Ibiam Chihurumnaya & Kshitij Shah as part the Dedicated Mentorship Program 2025.
I sincerely thank my mentors, the Sugar Labs, and C4GT community for their continuous guidance and support throughout this project.
It has been an incredibly rewarding summer working on this project, where I not only had the opportunity to contribute to open-source development but also to learn, grow, and collaborate with an inspiring community. This experience has motivated me to carry forward the spirit of open source, to continue contributing, and to build tools and software that make technology more accessible and impactful for everyone.