I'm a Applied AI Engineer and full-stack developer with hands-on production experience building real-time Voice AI systems, multi-channel conversational agents, and LLM-powered backend pipelines.

My core strength is architecting end-to-end AI systems — from low-latency voice-to-voice (V2V) streaming pipelines using LiveKit and WebRTC, to LangChain-orchestrated chatbot services with MCP-based enterprise tool integrations. On the full-stack side, I build scalable backend services with Python/FastAPI and frontend interfaces with React.

I specialize in agentic AI systems, RAG pipelines, multi-provider LLM routing, and real-time communication stacks — and I've shipped these in production across web, WhatsApp, and telephony channels.

• 🎙️ Currently building real-time Voice AI and MCP-based AI assistant integrations at EaseMyTrip
• 🧠 Deep focus on STT → LLM → TTS pipelines, multi-agent orchestration, and low-latency systems
• 🛠️ Working daily with LangChain, LangGraph, LiveKit, FastAPI, and multi-provider LLM routing
• ✍️ I write about real-time audio AI, LLM orchestration, and system-level optimization on X
• 🌱 Currently exploring MCP / UCP protocol design and cross-platform AI assistant widget delivery

🎧  Real-time Voice AI            STT → LLM → TTS pipelines, V2V systems
📞  Telephony & WebRTC            LiveKit, Twilio, Exotel, VoIP protocol engineering
🤖  Multi-Agent Orchestration     LangChain, LangGraph, agent handoff architectures
⚡  Low-Latency Inference          Token-aware model routing, cost-optimized LLM switching
🏗️  Scalable Backend Systems      Async-first architecture, FastAPI, Docker, PostgreSQL
🔍  Applied RAG Systems           Vector search, Pinecone, FAISS, grounded Q&A
🔌  MCP / Agentic Tooling         Model Context Protocol integrations for ChatGPT & Claude

Aug 2025 – Present · Gurugram, Haryana, India

Architecting and delivering production-grade Conversational AI and Voice AI systems serving customers across Web, WhatsApp, and Telephony channels.

• Engineered and launched EaseMyTrip's MCP integration for ChatGPT and Claude, transforming internal travel APIs into production-grade MCP tools with conversational tool-calling support for bookings, itinerary management, and customer assistance directly inside ChatGPT and Claude.
• Designed and implemented a real-time voice streaming pipeline using LiveKit (WebRTC), Silero VAD, Groq Whisper STT, and multi-provider TTS (ElevenLabs, Azure, Google Wavenet, Sarvam) — achieving sub-900ms round-trip latency through chunked audio streaming, adaptive buffering, and preemptive interruption handling.
• Engineered the WhatsApp Travel Bot, integrating Meta Webhook APIs, Redis-backed session management, and the core chatbot service for seamless multi-channel AI-driven travel assistance.
• Developed the EMT Chatbot core service with a dual-chain LangChain architecture: a tool-calling chain for intent classification and dynamic tool invocation, paired with a dedicated LLM summarization chain for structured, channel-aware responses.
• Contributed to the EMT Tools Ecosystem — an MCP-based tool-calling framework powering flight search, hotel lookup, train/bus booking, pre/post-booking flows, and OTP-based authentication.
• Applied DeepFilterNet for real-time noise suppression and conducted R&D across OpenAI, Gemini, Anthropic, Deepgram, Google, Sarvam, and Groq voice models — cutting per-call operational cost and improving inference throughput via token-aware routing.
• Practiced async-first development, containerized all services with Docker, and collaborated with DevOps on production deployment and scalability.

Gurugram, India

• Designed and optimized PostgreSQL query pipelines for data extraction workflows serving 20+ enterprise clients with 100,000+ records, improving reporting efficiency by 30%.
• Built RESTful APIs and contributed to schema optimization, reducing query execution time across core reporting modules.
• Explored LLM integration patterns and contributed to backend modules supporting AI-assisted features within the platform.

2024 – 2025 · Punjab, India

• Led technical workshops and mentored 100+ club members on AI/ML, full-stack development, and competitive programming.
• Organized hackathons, coding sprints, and project showcases, fostering a culture of applied engineering and product thinking.
• Guided juniors in building production-ready AI and web projects, strengthening the club's presence at university and national levels.

A fully automated voice-to-voice citizen complaint registration system — users interact entirely through natural speech over a live call, with no human agent involved. The pipeline ingests live audio, transcribes via ASR, reasons over an LLM, and delivers a synthesized spoken response end-to-end.

• Real-time audio stack using GPT-4o-mini-realtime-preview streaming for telephony (Twilio/Exotel), with custom G.711 µ-law converters, jitter buffer tuning, RTP optimization, and FFmpeg preprocessing.
• POC across three voice pipeline architectures: OpenAI GPT-4o Realtime (PCM16 WebSocket), Gemini Native Audio Thinking Model, and a custom STT–LLM–TTS micro-pipeline (Whisper → Groq GPT-OSS → Google Wavenet) — with token-aware routing and latency-based model swapping.
• Silero VAD with server-side threshold tuning + DeepFilterNet noise suppression, eliminating false-trigger interruptions in noisy call environments.
• Multi-stage NLP extraction for citizen intent, complaint category, and brand details — persisted to Redis with async PostgreSQL for durable storage.

A four-microservice LLM evaluation platform: an Orchestrator (API gateway, JWT auth, org/project management, job orchestration), a Model Runner (FastAPI + LiteLLM multi-provider proxy), an Eval Engine (automated metric scoring), and MongoDB for metadata persistence.

• Provider-agnostic inference proxy via LiteLLM, normalizing OpenAI, Anthropic, Gemini, Mistral, and Azure into a single OpenAI-compatible interface with per-provider token and latency tracking.
• WebSocket-based Realtime Service for live voice evaluation, integrating OpenAI Realtime API and Gemini BidiGenerateContent for audio-in/audio-out benchmarking.
• Eval Engine supporting statistical metrics, predefined LLM-as-judge metrics (relevance, groundedness, coherence, fluency), and fully custom criteria with configurable judge model, threshold, and per-metric weighting.
• Multimodal evaluation pipelines for AI-generated image/video — scoring physics plausibility, anatomical correctness, semantic adherence, temporal consistency, and aesthetic quality via vision-capable LLM evaluators.
• Progressive batch evaluation system with per-row score persistence to GCS-hosted CSV datasets, enabling real-time progress polling and partial recovery on mid-job failures.

Led the integration of EaseMyTrip's travel services into ChatGPT and Claude via the Model Context Protocol (MCP).

• Exposed enterprise travel APIs as production-ready MCP tools for flight search, hotel discovery, itinerary planning, booking assistance, and post-booking operations through natural conversation.
• Architected a cross-platform widget delivery framework dynamically serving Claude-optimized experiences from a unified backend, API layer, and business logic.
• Designed secure MCP tool integration, parameter extraction pipelines, and channel-aware conversational workflows for enterprise-grade reliability.

An AI SaaS platform automating end-to-end video creation from text/document inputs — covering LLM-powered summarization, multilingual voiceover synthesis, sign-language video generation, and AR/VR scene integration for accessibility in education and corporate training.

• AuraBot — a knowledge-base-grounded RAG assistant for contextual Q&A over uploaded course/training material.
• Analytics dashboard tracking viewer engagement, quiz completion, comprehension scores, and retention — personalizing content delivery per learning profile.

A production-grade multi-agent system orchestrating five specialized agents with asynchronous coordination and persistent state management.

• Built with LangChain, LangGraph, and FastAPI; enterprise-grade practices including structured logging, exception handling, CI/CD, UV package management, and Husky git hooks.
• Multi-provider LLM integration (OpenAI, Groq) with concurrent asyncio workflows — ~95% uptime.

An enterprise-grade AI automation platform with autonomous micro-agents for email automation, voice-based FAQ/data collection, WhatsApp outreach, and analytics dashboards.

• Built on LLMs, FastAPI, Docker, Pinecone, and LangChain orchestration with semantic search and real-time telephonic interactions.
• Delivered ~50% operational efficiency gains and 40% higher stakeholder engagement.

Languages

AI / ML & Agentic Systems

Voice & Real-Time Systems

Backend & Infrastructure

Frontend

Tooling & Platforms

🏆 Top 22 National Finalist & Social Buzz Winner — Bajaj Finserv HackRx 5.0 🏆 Top 50 Finalist — Cars24 Token'26 (hosted by OpenAI, AWS & ElevenLabs) 🏆 Winner — Aditya Birla Group's Synaptix Hackathon '25 (Delhi Technological University) 🏆 Selected — GitHub Field Day '24 at Microsoft, Gurgaon 🏆 First Runner-Up — FusionFest Hackathon, Chitkara University 🏆 Finalist & Social Buzz Winner — Vihaan 7.0, Delhi Technological University

Bachelor of Computer Applications — Information Technology Chitkara University, Punjab, India · CGPA: 9.12

I'm interested in roles and collaborations involving Applied AI, Voice AI, LLM Systems, and scalable backend engineering — and I'm open to research-driven engineering problems and early-stage product work.