Jul 2025 – Present · 11 mos · Pune District, Maharashtra, India · Hybrid

• Reduced compute cost to $0.064 per hour of audio via GPU batching, parallel chunk inference, and pipeline optimization.
• ​Improved multilingual WER by 20% on customer datasets (English/French/Latin) and 12% for Mandarin/Japanese.
• ​Maintained p95 inference latency under 2.8s per 30s chunk in production while processing 150+ hours/day of multilingual audio.
• ​Architected cloud-native systems on AWS EKS with queue-based ingestion, autoscaling workers, and SLO-driven monitoring.
• Architected and deployed a production-grade transcription and speaker diarization pipeline leveraging Faster-
• Whisper, PyAnnote, and NVIDIA GPU acceleration, improving accuracy and reducing latency.
• Implemented dynamic speaker diarization using PyAnnote, enabling accurate multi-speaker separation across long audio
• files.
• Replaced Silero VAD with ONNX-optimized GPU-based VAD, enabling real-time audio splitting and scaling efficiently
• for enterprise workloads.
• Designed parallelized, chunked inference pipelines with Multi-Processing and GPU memory management, boosting
• throughput by 3x.
• Integrated retrieval-augmented transcription workflows (RAG-style) to support multilingual-to-English translation us-
• ing Hugging Face, NLLB and Helsinki-Opus Mul to En model.
• Developed a robust audit and monitoring framework for transcription jobs, ensuring complete success/failure/skip track-
• ing across pipeline stages.
• Led offline-first model packaging initiative: centralized preloading of Faster-Whisper, PyAnnote into Docker/EFS for
• fully disconnected deployments.
• Improved Recall to 0.86 and Accuracy to 0.79 with Reduction Inference Time by using Parallel
• KV Cache Compute due to which the performance gain was from 1 request per 15 seconds to 5
• requests per 18 seconds without drop in accuracy.
• Established LLMOps and Observability Metrics for the GenAI Applications

Jul 2023 – Jun 2025 · 1 yr 11 mos · Pune District, Maharashtra, India · Hybrid

• Playing a pivotal role in replacing the existing transcription service with a new, optimized solution.
• Leading the transition to a high-
• Migrated SQL-compliant products to PostgreSQL, enhancing database efficiency.
• Implemented monitoring metrics logging, improving observability by 30%.
• Received a Spot Award for exceptional contributions to the project.