Nikita Verma

I build production-grade multilingual real-time ASR/STT systems, wiring streaming microphone and audio file flows into partial and final transcripts. At EXL, I architected WebSocket ASR servers with FastAPI and Uvicorn for chunk-wise PCM streaming, silence-based finalization, and JSON partial/final transcript events.

I’ve benchmarked ASR models across Parakeet, Riva, Whisper, Google STT, and Azure STT—measuring latency and transcript quality—and turned results into Excel-based evaluation reports with per-file model analysis. I also containerize GPU-accelerated inference services and troubleshoot real deployment issues across CUDA, NeMo imports, WebRTC VAD, torch/CUDA mismatches, and build failures.

For serving and operations, I’ve deployed real-time ASR on GCP (Compute Engine, Artifact Registry, Cloud Run) and AWS EC2, handling VM provisioning, Docker registry pushes, and endpoint health monitoring. I’ve also evaluated Triton, NIM, Riva, and vLLM for speech AI model serving to choose production scalability paths.

Previously, at SakhaTech Information Systems Pvt. Ltd., I orchestrated AI/ML/LLM pipelines and delivered ASR/TTS models for real-time translation. I built a RASA chatbot and then upgraded it with LLaMA + RAG for an intelligent virtual assistant, including fine-tuning GPT-3/GPT-4 on proprietary datasets.