Spanish Tutor Agent
Conversational Language Learning
A conversational AI agent purpose-built for Spanish language learning. The agent conducts real-time bilingual conversations, provides inline grammar corrections with explanations, adapts vocabulary and complexity to the learner's proficiency level, and maintains long-term progress tracking across sessions. Voice-first design with Whisper for speech recognition and ElevenLabs for natural Spanish speech synthesis. Curriculum-aware — follows structured lesson progressions while allowing freeform conversation practice.
What Was Built
Voice Conversation Engine
WebRTC audio capture, Whisper transcription, Claude for response generation, ElevenLabs for native-quality Spanish TTS. Sub-2-second round-trip latency.
Adaptive Difficulty System
proficiency model tracks grammar accuracy, vocabulary range, and response complexity. Automatically adjusts conversation parameters per session.
Grammar Correction Pipeline
inline corrections with explanations in the learner's native language. Tracks recurring errors and generates targeted practice exercises.
Progress Tracking
session-over-session metrics: vocabulary acquired, grammar accuracy trends, conversation duration, topics covered. Exportable progress reports.
Curriculum Engine
structured lesson plans (A1 through C2) with grammar focus areas, vocabulary sets, and cultural context. Freeform mode available at any level.
Memory System
persistent learner profile with conversation history, mistake patterns, vocabulary gaps, and personal interests for contextual conversations.
How It's Built
FastAPI backend with WebSocket connections for real-time audio streaming
Whisper API for speech-to-text with Spanish language optimization
Claude API with structured system prompts for pedagogical conversation
ElevenLabs API with multilingual v2 model for native Spanish speech
Supabase for user profiles, progress data, and session history
Redis for session state and real-time conversation context caching
Next.js 15 frontend with WebRTC audio capture and waveform visualization
Edge deployment on Vercel + Railway for low-latency audio processing
Technical Stack
Results & Impact
Voice-first Spanish learning with sub-2-second conversation latency
Learners progress through structured A1 to C2 curriculum at their own pace
Grammar accuracy tracking shows measurable improvement across sessions
Persistent memory creates personalized, contextual conversations