LingoLearn - Adaptive AI Language Tutor with Amazon Nova

An adaptive AI tutor that creates personalized lessons and helps learners actually speak and write German not just memorize it.

Comment

Inspiration

Most language learning apps treat every learner the same. Whether you're a beginner or intermediate learner, you are often forced through the same static content, repetitive drills, and memorization-heavy exercises.

We wanted to build something closer to a real tutor — a system that understands the learner, teaches concepts before testing them, and adapts based on performance.

That idea became LingoLearn.

What it does

LingoLearn is an AI-powered adaptive language tutor that generates personalized German lessons using Amazon Nova.

Instead of static lessons, the system:

  • Dynamically generates lesson content using LLMs
  • Evaluates learner responses across reading, listening, writing, and speaking
  • Updates a learner state to track progress and weaknesses
  • Adapts future lessons automatically

Each lesson follows a teach-first approach:

  1. Learn (vocabulary, grammar, dialogue)
  2. Practice (unscored, retryable)
  3. Assessment (scored, structured evaluation)
  4. Feedback + progress tracking

How we built it

We designed the system as an agent-based architecture powered by Amazon Nova:

  • Planner Agent → selects topic, grammar, difficulty
  • Content Generation Agent → creates full lesson packages
  • Evaluation Agent → analyzes learner responses

- Memory Agent → updates learner state and adapts future lessons

Tech Stack

  • Frontend: Flutter
  • Backend: FastAPI (Python)
  • AI/LLM: Amazon Nova (Lite + Pro + Canvas)
  • Storage: AWS S3 (images, audio)
  • Architecture: Modular services + schema-driven generation

We also implemented:

  • structured lesson schemas using Pydantic
  • section-wise generation to improve reliability
  • presigned URLs for secure media delivery

Challenges we ran into

The biggest challenge was generating large structured lesson content reliably using LLMs.

Initially, we attempted to generate the entire lesson in a single response, which caused instability. We solved this by:

  • breaking lesson generation into smaller components
  • assigning different models for planning vs content generation
  • enforcing schema validation

We also faced challenges with:

  • S3 media handling (audio/image uploads and retrieval)
  • Bedrock model configuration and region compatibility
  • ensuring smooth frontend-backend integration under time constraints

Accomplishments that we're proud of

  • We’re most proud of turning a complex idea into a working, end-to-end product within a short hackathon timeline.

  • LingoLearn is not just a UI prototype or a simple chatbot — it is a fully functional system that dynamically generates lessons, evaluates user performance, and adapts future learning paths in real time.

  • One of our biggest achievements was designing and implementing an agent-based architecture using Amazon Nova. We separated planning, content generation, evaluation, and learner state management into distinct components, making the system modular and scalable.

  • We also built a teach-first learning experience, shifting away from traditional apps that immediately test users. By introducing structured learning, guided practice, and protected assessments, we created a more natural and supportive learning flow.

  • Another key accomplishment was solving structured lesson generation reliability. Instead of relying on a single large LLM response, we engineered a modular pipeline with schema validation to ensure consistency and quality.

On the engineering side, we successfully integrated:

  • Flutter frontend with an intuitive user experience
  • FastAPI backend with modular service design
  • AWS S3 for media storage and delivery

  • Amazon Nova models for intelligent lesson generation

  • Despite time constraints and challenges with model reliability, API integration, and media handling, we delivered a polished and functional product.

  • Most importantly, we built something that feels less like a tool and more like a real AI tutor — and that’s something we’re genuinely proud of

    What we learned

    This project taught us:

  • how to design agent-based LLM systems

  • how to combine structured data + generative AI

  • how to build end-to-end AI products under tight timelines

  • and most importantly, how to move from "LLM demo" to a real usable product

What's next for LingoLearn - Adaptive AI Language Tutor with Amazon Nova

  • speech evaluation using audio embeddings
  • spaced repetition for vocabulary retention
  • multi-language support
  • reinforcement learning for lesson optimization

Built With

Share this project:

Updates

Submission history