DEV Community

Cover image for Gemma Mentor AI: From an Unfinished Prototype to a Real-Time Multi-Agent Learning Companion
Darlington Mbawike
Darlington Mbawike

Posted on

Gemma Mentor AI: From an Unfinished Prototype to a Real-Time Multi-Agent Learning Companion

GitHub “Finish-Up-A-Thon” Challenge Submission

This is a submission for the GitHub Finish-Up-A-Thon Challenge.

Gemma Mentor AI: From an Unfinished Prototype to a Real-Time Multi-Agent Learning Companion

What I Built

Gemma Mentor AI is a mobile-first AI tutoring platform designed to help learners understand complex topics through intelligent explanations, document analysis, personalized lessons, and interactive learning experiences.

The project originally started as a prototype built around a simple question:

Can AI become a true learning companion rather than just another chatbot?

The initial version demonstrated the concept, but many of the capabilities I envisioned were either incomplete, experimental, or not yet production-ready.

The GitHub Finish-Up-A-Thon Challenge gave me the perfect opportunity to revisit the project, fix long-standing issues, redesign major systems, and finally build the version I had imagined from the start.

Today, Gemma Mentor AI has evolved into a real-time multi-agent tutoring platform where specialized AI agents collaborate to teach, critique, explain, and refine educational content before it reaches the learner.


GitHub Repository

https://github.com/darchumsone-collab/gemma-mentor-ai

Video Walkthrough

Full Demo:

https://youtu.be/4MozenjT2OY

Quick Showcase:

https://youtube.com/shorts/HRxil56RIWg


Project Evolution

Before

The original version demonstrated the idea but contained several limitations:

  • Single-agent tutoring workflow
  • Basic AI explanations
  • Experimental document analysis
  • Limited personalization
  • Incomplete learning persistence
  • Various rendering and state-management issues

After

The finished version includes:

  • Multi-agent tutoring architecture
  • Teacher, Critic, and Explainer AI agents
  • Advanced document intelligence
  • Firebase-powered persistence
  • Learning Vault
  • Real-time tutoring workflows
  • Improved UI and stability
  • Scalable backend foundation

The Comeback Story

Where It Started

Like many side projects and hackathon builds, Gemma Mentor AI was created under significant time pressure.

The prototype proved that AI tutoring was possible, but several important systems remained unfinished.

Single-Agent Responses

The original architecture relied on a single AI response pipeline.

Although functional, responses sometimes:

  • Missed educational context
  • Lacked alternative explanations
  • Varied in depth and quality
  • Focused on answering rather than teaching

The system could generate responses, but it was not yet a true tutoring experience.

Document Intelligence Limitations

As testing expanded across different educational materials, weaknesses in the document analysis pipeline became clear.

Examples included:

  • Generic summaries from specialized content
  • Weak subject differentiation
  • Inconsistent concept extraction
  • Limited understanding of educational structure
  • Similar learning outputs from different documents

These issues reduced the effectiveness of personalized tutoring.

Rendering and Stability Issues

As the application grew, several Flutter rendering and lifecycle issues appeared throughout development.

Examples included:

LateInitializationError:

LateInitializationError:
Field '_controller' has not been initialized.

Render Overflow:

RenderFlex overflowed by 57 pixels on the bottom.

Null Safety Failures:

Null check operator used on a null value.

Lifecycle Errors:

setState() called after dispose()

Constraint Issues:

BoxConstraints forces an infinite height.

Viewport Rendering Problems:

A RenderViewport expected a child of type RenderSliver.

While each issue seemed small in isolation, together they affected stability and user experience.

State Synchronization Problems

Additional issues included:

  • Duplicate messages
  • Response ordering inconsistencies
  • Stale cached content
  • Loading deadlocks
  • Session synchronization problems

As more advanced AI capabilities were added, these issues became increasingly visible.


The Decision to Finish It Properly

Rather than continuing to patch individual problems, I decided to revisit the architecture itself.

The goal shifted from building an AI chatbot to building an AI tutoring system.

I wanted a platform capable of:

  • Teaching concepts
  • Adapting explanations
  • Encouraging understanding
  • Supporting long-term learning
  • Providing personalized educational experiences

This required a substantial redesign.


Building a Multi-Agent Tutoring System

The most significant upgrade was replacing the single-agent workflow with a collaborative multi-agent architecture.

Instead of generating a response from a single perspective, specialized AI agents now contribute to the final educational output.

Teacher Agent

Responsible for:

  • Lesson generation
  • Educational structure
  • Concept explanation
  • Learning guidance

Critic Agent

Responsible for:

  • Evaluating quality
  • Identifying gaps
  • Detecting weak explanations
  • Improving completeness

Explainer Agent

Responsible for:

  • Simplification
  • Analogies
  • Alternative explanations
  • Student-friendly teaching

Response Composer

Responsible for:

  • Merging outputs
  • Maintaining consistency
  • Improving readability
  • Delivering final responses

Workflow:

Student Question

Teacher Agent

Critic Agent

Explainer Agent

Response Composer

Student

The result is a significantly richer learning experience compared to traditional single-response AI systems.


Rebuilding Document Intelligence

One of the largest improvements involved redesigning how uploaded documents are analyzed.

Before

The system could process files but often produced generalized educational outputs.

Different documents frequently generated explanations that felt too similar.

After

The upgraded document intelligence system now focuses on understanding:

  • Subject domains
  • Topic hierarchies
  • Learning objectives
  • Key concepts
  • Educational context
  • Knowledge dependencies

Lessons are now generated based on the structure and intent of the source material rather than relying on generic educational templates.

This dramatically improves personalization and relevance.


Firebase-Powered Learning Infrastructure

To support future growth, I redesigned the backend architecture using Firebase.

Implemented services include:

  • Firebase Authentication
  • Cloud Firestore
  • Real-time synchronization
  • Session persistence
  • Learning history
  • Cloud-based storage systems

This provides a scalable foundation while preserving learning progress across sessions.


Learning Vault

One of the most meaningful additions is the Learning Vault.

Students can now:

  • Save lessons
  • Store explanations
  • Review previous learning sessions
  • Organize knowledge
  • Build a personal educational library

The platform evolved from a question-answer tool into a long-term learning environment.


Real-Time Tutoring Experience

The tutoring workflow was redesigned to feel more natural and educational.

Key improvements include:

  • Faster response generation
  • Better conversational flow
  • Improved context awareness
  • Enhanced continuity between sessions
  • More consistent educational structure

The result feels much closer to interacting with a real mentor.


Neural Learning Interface

The user experience received a substantial redesign.

The goals were:

  • Reduce cognitive overload
  • Improve readability
  • Prioritize educational content
  • Create a modern AI-native experience

The updated interface places learning first while maintaining a clean and responsive design.


Technical Improvements

A large portion of the project focused on engineering quality rather than simply adding features.

Fixed Rendering Issues

Resolved:

  • RenderFlex overflow errors
  • LateInitializationError exceptions
  • Null safety failures
  • Lifecycle management issues
  • Constraint conflicts
  • Viewport rendering mismatches

Fixed State Management Problems

Resolved:

  • Duplicate responses
  • Out-of-order messages
  • Stale session data
  • Synchronization inconsistencies
  • Loading deadlocks

Performance Optimizations

Implemented:

  • Efficient widget rebuild strategies
  • Better stream handling
  • Firebase caching
  • Lazy loading
  • Reduced unnecessary rendering

These improvements significantly improved reliability and responsiveness.


What Makes Gemma Mentor AI Different?

Many AI education tools focus on generating answers.

Gemma Mentor AI focuses on helping learners understand.

The platform is designed around educational reasoning rather than simple response generation.

Its purpose is to:

  • Explain concepts clearly
  • Adapt to learner needs
  • Reinforce understanding
  • Build knowledge progressively
  • Encourage deeper learning

Rather than acting as a single AI assistant, the platform simulates collaboration between multiple educational specialists.


My Experience with GitHub Copilot

GitHub Copilot played an important role throughout the revival of Gemma Mentor AI.

When I returned to the project, I wasn't starting from scratch. I was working with an evolving codebase that contained experimental features, unfinished components, architectural inconsistencies, and technical debt accumulated during rapid prototyping.

Copilot helped me move from experimentation to completion.

One area where Copilot provided significant value was Flutter development. It accelerated the creation of widgets, data models, service classes, and Firebase integrations, allowing me to spend more time improving the educational experience rather than writing repetitive boilerplate code.

As the project evolved into a multi-agent tutoring system, Copilot assisted with restructuring logic, generating implementation ideas, and helping refactor older code into cleaner and more maintainable components.

Copilot was also valuable during debugging. While resolving rendering issues, lifecycle problems, and state synchronization challenges, it often suggested fixes, alternative implementations, and code improvements that reduced development time.

Some of the areas where GitHub Copilot helped include:

  • Flutter UI development
  • Firebase integration
  • State management improvements
  • Multi-agent workflow implementation
  • Refactoring
  • Debugging
  • Data model generation
  • Performance optimization

Perhaps the biggest benefit was maintaining development momentum.

Instead of spending hours on repetitive implementation tasks, I could focus on educational workflows, tutoring strategies, and user experience improvements.

The Finish-Up-A-Thon Challenge is about finishing what you started, and GitHub Copilot helped me do exactly that.


Lessons Learned

One of the biggest lessons from this project is that unfinished projects still have potential.

Many projects do not fail because the idea is bad.

They remain unfinished because time runs out before the vision can be fully realized.

Returning to Gemma Mentor AI allowed me to revisit earlier decisions, solve long-standing issues, and transform a promising prototype into a much more capable learning platform.


What's Next?

Future plans include:

  • Adaptive learning paths
  • Voice tutoring
  • Multi-modal learning
  • Enhanced educational memory systems
  • Collaborative learning features
  • Advanced learning analytics

The foundation is now in place for continued growth.


Final Thoughts

Most unfinished projects do not fail because of bad ideas.

They fail because time runs out before the vision can be fully materialize.

Gemma Mentor AI began as an experimental tutoring prototype.

Through this challenge, I revisited old code, fixed architectural weaknesses, resolved rendering issues, redesigned major systems, and transformed the project into a significantly more capable learning platform.

The most rewarding part of this journey was not adding another feature.

It was proving that unfinished projects still deserve a second chance.

The GitHub Finish-Up-A-Thon Challenge encouraged me to finish what I started, and Gemma Mentor AI is stronger because of it.

Thank you for reading.

Top comments (0)