Project Codexa

Your personal AI tutor, no internet required.

Comment

Project Codexa

Inspiration

The inspiration for Project Codexa is deeply personal. It stems from witnessing firsthand how millions of bright, talented students in rural India are left behind—not due to a lack of ambition, but a lack of access.

The two greatest walls they face are:

  • The digital divide: no consistent internet
  • The language barrier: most high-quality technical content is in English

We were driven by a single question:
How can we empower a student to learn to code if they can't even understand the tutorial?

Project Codexa is our answer—a tool to shatter these barriers and prove that opportunity should not be dictated by a student's location or internet connection.

What it Does

Project Codexa is a complete, offline-first AI learning platform that transforms a basic laptop into a personal, multilingual tutor.

Key Features:

  • 100% Offline AI Tutor
    A multi-subject expert in Math, Sciences, English, and Coding that runs entirely on the user's machine.

  • Adaptive Bilingual Support
    Defaults to English but intelligently detects and switches to a student’s native Indian language (like Telugu or Malayalam) to provide explanations and generate quizzes in that language.

  • Socratic Teaching Method
    Not an answer machine—it guides with open-ended questions to foster critical thinking.

  • Personalized Dashboard & Local Database
    A SQLite-powered dashboard for students and parents to track progress, quiz accuracy, and streaks—all stored privately on the user’s machine.

  • Gamification Engine
    Awards badges (e.g., "Python Pioneer") and tracks daily learning streaks to keep students motivated.

  • Smart, Proactive Quizzing
    AI generates interactive quizzes after lessons and can create “mixed review” quizzes based on a student’s history.

  • Indian Syllabus-Tailored Content
    Explanations and examples align with the standard Indian school curriculum, making learning directly applicable.

  • Text-to-Speech (TTS)
    Accessibility feature that lets the tutor’s responses be read aloud at the click of a button.

How I Built It

  • AI Backend
    Used Ollama to run open-source LLMs (gpt-oss:20b, Mistral-7B) locally with GPU acceleration.

  • The Tutor’s Brain
    A sophisticated SYSTEM_PROMPT, designed with strict, mode-based rules that define tutor personality, bilingual support, Socratic style, and quiz formatting.

  • Local Database
    SQLite stores user profiles, quiz results, streaks, and gamification data.

  • Frontend Interface
    Built with Streamlit. Features a real-time chat tutor plus a separate data-driven dashboard.

  • Core Logic
    Written in Python, using regex parsing for AI responses and robust state management for conversation flow.

Challenges We Ran Into

  • Extreme Performance Bottlenecks
    Running a 20B parameter LLM on consumer hardware initially took 4+ minutes.
    Solved with GPU acceleration and a dual-model strategy (gpt-oss for power, Mistral for speed).

  • AI Unreliability
    Hallucinations and broken quiz formatting.
    Solved with strict, mode-based prompt engineering and dozens of refinements.

  • UI Bugs in Stateful App
    Streamlit issues caused disappearing messages and broken state.
    Solved by re-architecting state management.

Accomplishments I am Proud Of

  • Built a feature-complete, stable application that addresses a critical social problem.
  • Successfully implemented adaptive bilingual support to break India’s language barrier.
  • Designed a Socratic teaching AI that feels like a real teacher.
  • Overcame the immense technical challenge of running a powerful LLM offline on consumer-grade hardware.

What I Learned

  • A powerful model is just the beginning—prompt engineering is the real magic.
  • Hands-on experience in full-stack AI development (frontend, backend, and database).
  • Mastered real-world performance optimization for ML deployment.

What’s Next for Project Codexa

  • Mobile Version
    A standalone app using lightweight models like phi3:mini for broader accessibility.

  • Expanded Document Support
    Support for .pdf and .docx files in the Notes Summarizer.

  • Advanced Analytics
    Richer dashboards with deeper insights into learning patterns, strengths, weaknesses, and personalized study plans.

Built With

  • gpt-oss-20b
  • gpu-acceleration
  • mistral-7b
  • ollama
  • phi3-mini
  • prompt-engineering
  • python
  • pyttsx3
  • re
  • sqlite
  • streamlit
Share this project:

Updates