Skip to content
    CC
    Chetan Chhangani

    Deutsch Sprach Gym

    I built this independent personal project to address fragmented daily language-learning continuity across content, feedback, vocabulary, progress, and conversation history through a persistent-state workflow connecting Telegram, orchestration, AI generation and feedback, and GitHub-backed state.

    Independent BuildAI WorkflowTelegramHermes AgentOpenAI APIGitHub
    9 stages
    Daily workflow
    7 areas
    Persistent artifact model
    40 days
    B2 sprint

    The problem

    Generated learning material was easy to produce. The harder problem was maintaining continuity: remembering previous mistakes, creating targeted follow-up practice, preserving vocabulary and progress, and keeping the interaction lightweight enough for daily use.

    System architecture

    01TelegramDaily delivery and /deutsch response surface
    02Hermes AgentOrchestration, prompts, feedback, and mini-drills
    03OpenAI APILanguage generation and evaluation capability
    04GitHubPersistent state, metrics, vocabulary, and validation

    Telegram handles the daily interaction, Hermes coordinates generation and feedback, the OpenAI API provides language capability, and GitHub stores persistent state and learning artifacts outside transient conversation history.

    Daily learning loop

    1. 01Generate personalised material
    2. 02Deliver through Telegram
    3. 03Learner responds
    4. 04Provide immediate feedback
    5. 05Generate targeted mini-drill
    6. 06Capture vocabulary
    7. 07Update learning state
    8. 08Produce review summary
    9. 09Persist and validate artifacts

    Persistent learning state

    feedback/Corrections and feedback artifacts
    metrics/daily/Daily progress signals
    state/Current learning state
    word-bank/Captured vocabulary
    prompts/Reusable agent instructions
    scripts/Automation and validation
    templates/Consistent artifact structures

    These structured artifacts act as product memory, carrying forward the information required for future sessions without loading unrestricted conversation history on every run.

    What this project demonstrates

    • Scheduled daily execution
    • Telegram-based interaction
    • AI-generated feedback and targeted drills
    • Persistent GitHub-backed state
    • Bounded context
    • Structured feedback, metrics, vocabulary, prompts, scripts, and templates

    Reflection

    • Persistent state is part of the product architecture for agent-assisted systems, not just a memory feature.
    • A simple interaction surface can sit on top of a more complex orchestration and persistence layer.
    • Feedback is more useful when it changes the next practice action rather than ending with a correction.
    Back to Work