Clariox
Case Study Summary
Client Problem
My own startup. I identified that professionals needed AI-powered speech coaching but existing tools were generic and lacked real-time feedback capabilities.
My Role
Founder / CEO / CTO — built the strategy, chose the technology, negotiated tool partnerships, and led hands-on development
Result
Built and launched a production AI voice coaching platform with real-time speech analytics, multi-provider AI pipeline, and subscription billing.
Relevant For
AI/ML startups, EdTech, real-time communication tools
Clariox is an AI-powered speech coaching platform that helps professionals sharpen their communication skills through simulated meetings, real-time feedback, and detailed speech analytics. I led the project from initial proof of concept through MVP and into production release, owning product vision, system architecture, and hands-on engineering.
Role: CEO, CTO & Lead Architect
Stage: PoC → MVP → Production
Leadership & Architecture
- Founded and led the product from concept validation to production-ready MVP, making all critical technical and product decisions
- Designed the full-stack architecture: async Python backend (FastAPI) with a React/TypeScript SPA, connected via REST APIs and WebSocket channels for real-time coaching sessions
- Architected a multi-provider AI pipeline integrating OpenAI (GPT-4 for coaching, Whisper for transcription), Azure Cognitive Services (pronunciation & prosody assessment), and Deepgram (real-time STT), with a factory pattern enabling seamless provider switching
- Designed an event-driven WebSocket system powering three distinct real-time session types: live coaching dialogue, meeting simulation with AI-generated topics, and comprehensive speech workshops with pronunciation scoring
- Built a modular service layer with clean separation of concerns: transcription, text-to-speech, speech metrics, grammar analysis, and AI coaching, each independently testable and replaceable
- Implemented scalable storage architecture using Cloudflare R2 with chunked video uploads and presigned URLs for secure playback
- Established the data model and user isolation strategy on PostgreSQL (Neon), ensuring all data is scoped per user via Clerk JWT authentication
Key Technical Achievements
- Real-time speech analytics: live WPM, filler word detection, pause analysis, and volume tracking streamed to the frontend during practice sessions
- AI-powered feedback engine: automated scoring for clarity, structure, grammar, pronunciation accuracy, fluency, and completeness with actionable improvement suggestions
- Subscription billing: full Stripe integration with free/pro tiers, webhook-driven status management, and customer portal
- Production deployment on Google Cloud Run with CI/CD pipeline