            Voice AI & Real-Time Systems • 10 min read
          

Sub-Second Full-Duplex Conversational Voice AI: Orchestrating LiveKit WebRTC, Deepgram Nova-2, and Real-Time POS State Machines

Author: Shaheer Ul Hassan (Muhammad Shaheer) — Agentic AI & Systems Engineer

Engineering a sub-500ms full-duplex conversational voice agent in Python integrating LiveKit WebRTC, Silero VAD, Deepgram Nova-2, Groq LPU, and transactional POS state machines.

Canonical Resource: https://shaheerulhassan.com/blog/sub-second-full-duplex-voice-ai-livekit-webrtc

Related Technical Publications & System Notes

Computer Vision & Forensics • 12 min read
Engineering an NIST-Grade Error Level Analysis (ELA) Image Forensics Engine in Python →
Deep-dive into Error Level Analysis (ELA) for digital document verification, JPEG compression artifact analysis, and the open-source deeptrace-ela PyPI engine.
Agentic AI & RAG • 15 min read
Beyond Naive RAG: Adversarial Fact-Auditing & Citation Grounding with LangGraph →
Architectural breakdown of proposition decomposition, NLI entailment matrices, and autonomous cyclic routing nodes for zero-hallucination RAG pipelines.
Edge AI & Optimization • 10 min read
Zero-Dollar Edge AI: Quantizing Vision Transformers (CLIP ViT) on Free-Tier CPUs →
Production guide on running quantized Vision Transformers (CLIP ViT) and perceptual hash clustering in high-throughput, resource-constrained CPU environments.
MCP & Multi-Agent Systems • 14 min read
Solving the Stale Embeddings Crisis: Real-Time Hybrid RAG with Model Context Protocol (MCP) →
How Model Context Protocol (MCP) enables dynamic, event-driven vector updates to eliminate stale embeddings in production knowledge bases.
Desktop Systems & AI • 8 min read
Bounding Desktop RAM to 38MB: Architecting Viewport Virtualization for 500-Page Document Workstations in PySide6 →
How viewport virtualization and asynchronous QThreadPool rendering bound resident memory to visible +/- 2 pages, eliminating gigabyte memory leaks in desktop document workstations.