The MirrorVote project serves as a case study for building an AI-powered proof-of-concept (PoC) application. The system uses Google Gemini 2.5 Flash and Gemini 2.5 Flash-image models to analyze user outfits and provide recommendations. The architecture relies on Supabase for authentication, PostgreSQL database management, and Edge Functions, while Nginx on a VPS acts as a reverse proxy.
Development was conducted using Claude Code with source code hosted on GitHub and deployments managed via GitHub Actions. The infrastructure includes integration with Google OAuth for user login and Yookassa for payment processing. Security is enforced through PostgreSQL Row Level Security (RLS) to prevent unauthorized data access, following OWASP API security guidelines.
Operational costs for the service are driven by LLM API usage. Processing six images costs approximately $0.25, with an additional $0.05 for analysis, totaling $0.30 per session. The author notes that current operational expenses exceed the revenue model of selling five sessions for 100 ₽, highlighting the need for architectural optimization in production.
The project emphasizes that while AI accelerates coding, human engineering remains essential for system stability, security, and cost management. Key tasks include implementing idempotent payment webhooks, configuring automated TLS certificate renewals via Let's Encrypt, and setting up observability through tools like UptimeRobot to monitor service availability and API expenses.
