15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-631c4730b83c8223.js"],"Posts"] 16:T547,Hi HN,

Built a real-time voice AI agent console for a YC W25 startup assessment (Freya Voice). Focus was on production-ready implementation with minimal latency.

GitHub: https://github.com/05sanjaykumar/Freya-Voice-YC25-Assessment

Key specs: - 133ms average latency (voice input → AI response → audio output) - LiveKit for WebRTC streaming - Next.js frontend + Python FastAPI backend - Multi-stage Docker deployment - Full observability and session management - In-memory storage for speed (design trade-off for assessment scope)

Tech stack: - Frontend: Next.js, TypeScript, LiveKit client SDK - Backend: FastAPI, LiveKit server SDK, OpenAI - Infrastructure: Docker multi-stage builds, production configs

Design decisions I made: - Voice-first interface (no text fallback) to match real-world use case - In-memory session storage (speed over persistence for MVP) - LiveKit over WebSocket (proven real-time infrastructure) - Concurrent audio processing to hit latency targets

Built in ~1 week for the assessment. Didn't land the role (extremely competitive) but learned a ton about real-time systems and WebRTC optimization.

Happy to discuss the latency optimization techniques or design trade-offs!17:T4d7,I've noticed that many blog posts about papers (which are in PDF) use raster screenshots (JPG, PNG, etc.) of key diagrams, which makes it hard to zoom in on the details (e.g., on mobile) because they pixelate and turn blurry, unless the screenshot was taken at a much higher resolution than the space on the page, which enables some zooming in without much degradation.

I wondered why they don't use SVG (since the original source in the PDF is usually a vector diagram), and when I went to tr

Scraping Spotify Is EZ