AcuLearn — AI-Powered Course Advisor
AI-powered Telegram course advisor that interviews learners, assesses knowledge and goals, recommends a personalized module sequence, preserves context, and escalates unresolved cases to human support.
Problem
A static course catalog could not adapt to a learner’s existing knowledge, goals, or gaps. The experience needed to recommend only the relevant parts of a learning path.
Users
Prospective product-management learners in a Bozhan product competition.
My Role
Two-person product team; product management, conversation/prompt design, Telegram + AI integration, with technical implementation by Alireza.
Key Decisions
- Interview for both current knowledge and learning goal before recommending.
- Recommend course modules and sequencing rather than treat the experience as a quiz.
- Use a human fallback when the product cannot resolve the case safely.
Delivery & Technical Approach
- Implemented the Telegram and interview flows with PHP, webhooks, command routing, deep links, and MySQL/Medoo state.
- Integrated GPT through Metis, maintained conversation context, and connected the AI flow to a knowledge-base/RAG architecture.
Outcome / Impact
- Approximately 500 users started the flow.
- Approximately 300 completed the full AI-guided interview.
- Placed third in a product competition organized by Bozhan.
Constraints & Boundary
- Competition timeframe and prototype maturity.
- The public repository documents the Telegram gateway/orchestration layer; the AI interview implementation lived partly in a separate repository.
- No commercial-scale or production-traction claim.
Scope note: Competition prototype, not a commercial production system. The ~300 figure refers to full AI-guided interview completions.
Reflection & Next
- Conversation design, recommendation logic, and fallback behavior matter as much as the underlying model.
- Completion metrics are only useful when the flow being completed is defined precisely.
Next: Validate recommendation usefulness after the interview, not only interview completion, and strengthen retrieval/evaluation evidence for the knowledge-base layer.