𝗛𝗲𝗮𝗱𝗹𝗶𝗻𝗲: Most Voice AI agents still feel like glorified walkie-talkies. So I built one that actually listens like a human. Most voice bots today follow a rigid, linear loop: You speak ➔ Wait ➔ Bot speaks. If you dare interrupt them mid-sentence? They either completely ignore you or their context buffer gets totally mangled. 𝗧𝗼 𝘀𝗼𝗹𝘃𝗲 𝘁𝗵𝗶𝘀, I engineered Astra Duplex Agent — a sub-50ms, ultra-low latency, full-duplex conversational Voice AI built to handle natural human interruptions seamlessly. Here is what went into building the architecture under the hood: True Full-Duplex WebSockets: Built on FastAPI, allowing real-time bi-directional audio streaming instead of turn-based HTTP requests. Edge Voice Activity Detection (VAD): Deployed Silero VAD using ONNX Runtime to catch user barge-ins at the edge with near-zero overhead. LangGraph State Machine: Instead of a simple monolithic script, the conversation logic is modeled as a state graph, giving precise control over execution flow. Partial State Tracking & Memory Persistence: This was the trickiest part. When an interruption happens, Astra doesn't just stop audio playback — an async kill-switch halts Groq's LLaMA-3.1 mid-token, calculates the exact partial sentence actually spoken by the TTS, and saves only that partial context to an Upstash Redis buffer. Sub-50ms Inference & Streaming: Leveraged Groq LPU (LLaMA-3.1 + Whisper Large v3) paired with ElevenLabs Turbo v2.5 streaming. The result? An agent that you can interrupt mid-thought, change topics with on the fly, and ask follow-up questions without it losing historical context. Containerized with Docker and live on Render & Vercel. Check out the full architecture & live demo below! https://lnkd.in/eC6yak3i https://lnkd.in/eN38HGZD #VoiceAI #GenerativeAI #LangGraph #SystemDesign #FastAPI #MachineLearning #WebSockets #Python #AIENGINEERING

Interviu — Real-Time AI Mock Interview Platform

Built a voice-based AI mock interview platform solo — React and Node.js, with a WebSocket server bridging live voice conversations through the Gemini Flash Live API. Handled bidirectional audio streaming with 400–500ms observed latency, and built Redis-backed session recovery so a dropped connection resumes instead of losing the whole interview. Integrated Claude to turn raw transcripts into structured, context-aware feedback, and built the auth layer — JWT issuance, validation, protected routes — from scratch.