Headline

ML Engineer · LLM Inference · Voice AI · Multi-Agent Systems · Fine-Tuning · POC → Production at Scale · IIT Kharagpur

About

4+ years building and shipping AI systems in production at JustDial — not demos, not notebooks. I work across the full AI stack: LLM inference on H100 clusters, voice bots handling live customer calls, agentic pipelines, classic ML, and computer vision — all in production at scale. ━━━ What I've shipped ━━━ → LLM inference: Deployed Qwen3-32B FP8, Gemma-4-31B, Qwen3.5-35B on H100 SXM using vLLM. Pushed prefix cache hit rate from 5% → 80% by refactoring system prompts — 10× reduction in prefill compute. Debugged FlashInfer sm90a CCCL kernel mismatch on CUDA 12.4. → Voice AI: Architected LiveKit voice bots for JustDial customer calls in Hinglish. Deepgram nova-3 STT + ElevenLabs TTS + Groq LLM. TTFT < 1.2–1.5s end-to-end. → AI Agents: Built CS and HR chatbots with Google ADK (multi-turn, intent detection, RAG). 13-stage image scrutiny pipeline for listing moderation with Evidently drift monitoring. → Computer Vision: ViT/CLIP classification → 90% reduction in manual image tagging. YOLO + PaddleOCR KYC pipeline → 80% auto-approval rate. Processing time: hours → seconds. → TTS Fine-tuning: Fine-tuned OmniVoice on Hindi studio audio (ai4bharat/Rasa). UTMOS > 3.5, speaker-sim > 0.75. FastAPI inference server at < 80ms latency. ━━━ Core stack ━━━ LLM Serving: vLLM, TGI, FlashInfer, PagedAttention, FP8/AWQ/GPTQ, Prefix Caching, Tensor Parallel Voice AI: LiveKit, Deepgram, ElevenLabs, WebRTC, SIP, OmniVoice, F5-TTS Agents & RAG: Google ADK, LangChain, LlamaIndex, FAISS, ChromaDB, MCP CV & ML: ViT, CLIP, YOLO, PaddleOCR, Qwen-VL Infra: Docker, Kubernetes, Prometheus, Grafana, MLflow, Evidently, Redis, FastAPI Hardware: H100 SXM, A100, NVLink, CUDA Profiling IIT Kharagpur CS, 2022. Open to senior AI/ML roles and production AI consulting. 📬 phvk1611@gmail.com | Portfolio: phvk1611.github.io/portfolio1

Location

Greater Bengaluru Area

AI stack

hugging face

langchain

langfuse

livekit

Google ADK

vLLM

Work experience

Machine Learning Engineer at Justdial, May 2022 - Present · 4 yrs 6 mos, Bengaluru, Building production AI systems — from GPU kernel debugging to agentic pipelines. Deployed vLLM inference for Qwen3-32B FP8, Gemma-4-31B, and Qwen3.5-35B-A3B on H100 SXM; tuned to >2k tok/s at P95 < 2s. Isolated dynamic variables in system prompts — pushed prefix cache hit rate from 5% → 80%, cutting redundant prefill compute by 10×. Built agentic Customer Support and HR chatbots using Google ADK — multi-turn context, intent detection, RAG over internal docs. Architected LiveKit voice bot platform for JustDial business-owner calls; TTFT < 1.2–1.5s end-to-end. Created 13-stage image scrutiny pipeline (NSFW, blur, watermark, OCR-logo, duplicate) with Prometheus monitoring and Evidently drift detection. ViT/CLIP semantic classification: 90% reduction in manual image tagging across hotel, salon, healthcare verticals. YOLO+PaddleOCR KYC automation (Aadhaar, PAN, Passport): 80% auto-approval rate, processing time hours → seconds.

Summer Internship at Justdial, Apr 2021 - Jul 2021 · 4 mos, Bengaluru, Karnataka, India

Education

UG, Computer Science and engineering, Indian Institute of Technology, Kharagpur

Intermediate, MPC, Narayana Junior College, Nellore

Public links

medium: https://medium.com/@phvk1611