Nikita Verma

Built production-grade multilingual (English + Spanish) real-time ASR/STT systems using NVIDIA Parakeet-TDT-0.6B-v3, Parakeet-CTC-0.6B-ES, Parakeet-RNNT-1.1B, Nemotron Speech Streaming, Whisper Large-v3 Turbo, Google STT v2, and Azure Speech Services supporting live microphone and audio file streaming with partial and final transcript generation. Architected WebSocket ASR servers with FastAPI + Uvicorn featuring chunk-wise PCM streaming, silence-based finalization, manual flush triggers, and JSON partial/final transcript events; tuned VAD systems (WebRTC, Adaptive Energy, Silero) with silence thresholds (200ms-900ms) and chunk durations (10/20/30ms) for near real-time delivery without premature cutoffs. Conducted structured ASR model benchmarking across Parakeet-TDT/CTC-ES/RNNT, Riva, Whisper, Google STT, and Azure STT evaluating TTFT, TTFB, WER, first/average latency, context-switching score, and final transcript quality; generated detailed Excel-based evaluation reports with per-file model analysis. Containerized GPU-accelerated ASR inference services using Docker, CUDA 12.x (nvidia/cuda runtime), NIM, Whisper, Google STT, NIM deployments, and optimized Dockerfiles; resolved critical issues including NeMo import crashes, torch/CUDA mismatches, segfaults, PyAudio/ffmpeg build failures, and pydantic conflicts for high-availability production deployment. Deployed and operated real-time ASR services on GCP (Compute Engine, Artifact Registry, Cloud Run) and AWS EC2 managing VM provisioning, Docker registry pushes, port/firewall rules, SSH tunneling, and WebSocket/gRPC remote endpoint validation and health monitoring. Evaluated and compared NVIDIA Triton Inference Server, NIM, Riva, and vLLM for speech AI model serving assessing GPU utilization, streaming endpoint behavior, supported model discovery, and deployment strategies to determine optimal production scalability and inference performance.

Designed, built, and maintained ML/statistical models to optimize marketing performance across audience targeting, channels, campaigns, and customer journeys. Specialized in Marketing Mix Modeling (MMM), LTV prediction, lift estimation, and causal inference to inform strategic decisions. Designed and deployed a financial IVA chatbot in Python using Amazon Lex, Lambda, and Bedrock to support real-time user queries on Cash Out, Early Pay with ongoing integration of Amazon Polly and Connect for voice support. Performed large-scale data processing and exploratory analysis using Python, SQL, and Spark to uncover actionable insights and evaluate marketing performance metrics.

Orchestrated AI/ML/LLM models, improving automation by 30% using NLP and SciPy, and leveraging image recognition for enhanced insights. Developed and deployed ASR and TTS machine learning models using PyTorch, TensorFlow, and NumPy for real-time translation, integrated Face and OCR models for automate identification and data extraction, and enhanced customer satisfaction by 20% through AI solution integration. Built a traditional RASA-based conversational chatbot and later integrated LLaMA and RAG to transform it into an intelligent virtual assistant(IVA) for intelligent, real-time interactions. Deployed, customized and fine-tuned GPT-3 and GPT-4 on proprietary datasets, which elevated low performing model response accuracy and efficiency by over 40%. Specialized in supervised learning and classification algorithms enhancing model accuracy to 90% with dynamic retraining, data augmentation, and hyperparameter tuning. Managed MySQL database systems and designed real-time data pipelines using Apache Kafka and Pyspark. Skilled in coding with Python, Scala, and Shell Scripting, using and proficient in project management using Git/GitLab.