I Canceled My French Tutor and Built an LLM Tool That Does It Better
9 hours ago
- #language learning
- #spaced repetition
- #LLM tool
- The author canceled a $200/month French tutor due to slow progress and poor retention, despite quality instruction.
- Built a custom LLM tool with two main components: a knowledge base tutor for grammar and a speech tutor called Causons for conversation practice.
- The grammar tutor uses spaced repetition (SM-2 algorithm) via Claude to track topics, confidence, errors, and review intervals, adapting to the user's forgetting curve.
- Speech practice app uses Groq for speech-to-text (Whisper-large-v3), GPT-4o-mini for chat, and OpenAI TTS for text-to-speech, costing about $0.04 per 15-minute session.
- Key innovations include precise error tracking (e.g., specific grammatical mistakes), dynamic prompts separating stable policies from volatile data, and streaming audio to reduce latency.
- The tool improves retention by reviewing topics just before forgetting, offers flexible practice anytime, and reduces costs significantly compared to human tutoring.
- Lessons learned: switching from Llama to GPT-4o-mini fixed hallucinated grammar corrections; security (API keys in localStorage) is fine for personal use but not for broader deployment.