Vishal Deep Verma

About Vishal Deep Verma

AI Voice & Video Agents only matter if they work in real-world conditions.I build production-grade systems that handle realtime calls, process speech (STT/TTS), and route conversations intelligently — with reliability as a core requirement, not an afterthought.My focus is on:Realtime voice & video infrastructureIntelligent call routing and automationLow-latency system designZero-downtime, fault-tolerant architecturesThis means designing for edge cases, handling failures gracefully, and ensuring systems perform consistently under load.Tech stack:Languages: Node.js, TypeScript, PythonAI/LLMs: OpenAI, LangChain, vector databasesRealtime & Voice: WebRTC, Twilio, streaming pipelinesBackend & Infra: AWS, Docker, Redis, message queuesInterested in building AI systems that operate beyond demos — where performance, scale, and reliability actually matter.

Experience

1. Principal Ai Engineer Ai Voice, Video & Realtime Systems

   Iwish

   Dec 2024 — Present

   Built and delivered production-grade AI voice systems with realistic concurrency, low latency, and scalable architecture.Product 1: AI Contact Center Platform (Multi-Industry, US Healthcare)Supports 200–800 concurrent calls per cluster (horizontal autoscaling)Latency:~500–1200 ms end-to-end (streaming STT → LLM → TTS)First response time:~300–600 ms (streaming partial responses)Multi-tenant omnichannel (voice/SMS/video/chat) with strict isolationReal-time voice pipeline: LiveKit + Deepgram + Twilio + Silero VADRAG pipelines (pgvector + embeddings) for tenant-scoped knowledgeYAML-driven workflow engine (routing, escalation, scheduling)CRM integrations via abstraction layer (Salesforce, Dynamics, Zoho, Vtiger)Backend: FastAPI + PostgreSQL (async, high I/O throughput)Agent orchestration: AutoGen / LLM tool-calling (GPT-4o class models)Infra: Dockerized microservices, queue-based load leveling, autoscalingTech: Python, FastAPI, PostgreSQL, pgvector, LLM tool-calling, LiveKit, Deepgram, Twilio, Docker, ReactProduct 2: Dylan AI (Automotive – Real-Time Voice AI)Handles 150–500 concurrent live voice sessions (scales via worker nodes)Latency:~400–900 ms streaming response timeCall setup:~1–2 sec (SIP/WebRTC)Built real-time streaming pipeline (LiveKit + Pipecat)Telephony: SIP, RTP, WebRTC with multi-provider abstractionMulti-agent orchestration, dynamic routing, mid-call handoffsRAG + prompt pipelines for grounded responsesReal-time transcription, summaries, sentiment, analyticsMicroservices: Python, Node.js, TypeScriptDeployment: AWS/Azure with autoscaling, health checks, failoverTech: Python, Node.js, TypeScript, LiveKit, Pipecat, Deepgram, WebRTC, LangChain, RAG, Docker, AWS, Azure

Education

• Dr. A.P.J. Abdul Kalam Technical University

  Bachelor's degree, Computer Programming

Skills

• Recruiting
• Business Strategy
• Strategy
• Resume Search
• IT Recruitment
• Team Management
• Microsoft Office
• Microsoft Word
• Powerpoint
• Microsoft Excel
• Sourcing
• Vendor Management
• Global Talent Acquisition
• Nxcam
• Microsoft Powerpoint
• Screening
• Autocad
• Catia
• Market Research
• Microsoft Outlook
• Leadership