Piyush Gambhir.

Software Engineer II · Voice AI, Real-Time Infrastructure & Agentic Engineering

I run production voice agents on LiveKit and contribute upstream to the LiveKit Agents SDK (eight merged PRs across the Python and Node frameworks). At Spyne I took a conversational AI platform from zero to production, then led its migration onto LiveKit. I build with agents too: spec-first delegation to coding agents with verification gates, shipping Go, Python, and TypeScript in the same week.

Experience

Software Engineer II

Spyne · Full-time · Gurugram, India

May 2026 to Present
  • Leading the migration of production voice agents from Vapi to LiveKit Agents, owning architecture, telephony and SIP integration, and rollout, and collaborating directly with LiveKit’s engineering team along the way.
  • Contributed upstream to the open-source LiveKit Agents SDK: eight merged PRs across the Python (livekit/agents) and Node (livekit/agents-js) frameworks. They expose OpenAI Responses API controls used in production and preserve the original caller ID on Twilio warm transfers.
  • Took voice-agent infrastructure cost from $0.26 to $0.10 per minute in our deployment, roughly 62%, alongside the LiveKit migration and the provider and pipeline tuning that followed.
  • Scaling Spyne’s production conversational AI stack to support increasing traffic, concurrency, and reliability requirements, and strengthening it through disaster recovery, failure handling, and production readiness work.

Software Engineer I

Spyne · Intern from Jan 2024, full-time from Jan 2025 · India

Jan 2024 to May 2026
  • Built Spyne’s production conversational AI stack from zero to one on Vapi, powering dealership lead handling, follow-ups, and customer conversations, then began laying the groundwork for the transition to LiveKit.
  • Improved agent responsiveness and conversation quality through prompt optimization, streaming pipelines, turn-taking improvements, and infrastructure tuning.
  • Built the core telephony layer across Twilio, Vonage, and Telnyx, including caller identity, CNAM lookup, STIR/SHAKEN support, banned-number filtering, and A2P 10DLC workflows.
  • Shipped full-stack systems across inventory ingestion, dealer websites and CMS, CRM workflows, lead management, and automated customer follow-ups.
  • Rebuilt my development workflow around coding agents: spec-first delegation, parallel runs, and verification gates before accepting a diff. Monthly commits across 15 production repositories went from 50 to 90 up to roughly 300.

Projects

Conversational Voice Agents

Provider-neutral control plane for production voice agents, with interchangeable LiveKit and Pipecat workers driven by one typed call profile.

Telephony Voice Simulator

Provider-neutral test platform for voice agents, AMD, IVRs, PBX routing, live carrier behavior, recordings, and automated grading.

Reckon + Operations CLI Suite

Read-only production incident workspace plus Go CLIs for Grafana, Elasticsearch, Jenkins, Jira, CubeAPM, and Nginx Proxy Manager.

Document OCR

Local-first OCR toolkit for passports and Indian KYC documents, distributed as Python and TypeScript libraries with HTTP and Docker interfaces.

Skills

Agentic Engineering

Claude CodeCodexSpec-first delegationMulti-agent orchestrationVerification loops

AI & Voice

LiveKit (Agents, SIP, Cloud)PipecatVapiLLMsRAG

Telephony

TwilioVonageTelnyxSIPWebRTC

Languages

TypeScriptPythonGoSQL

Backend & Data

Node.jsNestJSFastAPIPostgreSQLRedis

Web

Next.jsReactTailwind CSS

Infrastructure

AWSDockerKubernetesCI/CDObservability

Writing

Voice agents in production

Series

Six-part field guide at piyushgambhir.com/voice: the LiveKit migration, turn-taking, voicemail detection, cost per minute, caller identity, and the carrier layer.

Education

The NorthCap University

B.Tech (AIML), Computer Science

GPA: 8.55 / 10

2021 to 2025