Looking for a job

ayush kushwaha

@ayushkushwaha2

Message

ML Engineer at WESEE, Indian Navy, who served Hindi and English speech models and cut transcription latency by 35%.

India

Message

At WESEE, Indian Navy, I served English and Hindi text-to-speech models on NVIDIA Triton and built a multilingual speech-to-text and translation pipeline. Batched inference and beam-search tuning cut transcription latency by 35%.

I migrated production RAG vector search from ChromaDB to Milvus for 10M+ embeddings, reducing retrieval latency from 120ms to 67ms. I also work on ML inference and serving, including load testing and deployment.

In research projects, I built a SAR-to-optical translation pipeline and developed multi-GPU training with PyTorch DDP and AMP. I also built a multi-agent data analysis system and contributed a fix to Unsloth.

Experience

Work history, roles, and key accomplishments

Current

Assistant Engineer – ML/DL

Current

WESEE, Indian Navy (contracted via Softel Solutions Pvt. Ltd

Jun 2025 - Present (1 year 4 months)

Served English and Hindi text-to-speech on NVIDIA Triton Inference Server, sustaining ~4 req/s at 20 concurrent clients with zero failures. Built a multilingual speech-to-text and translation pipeline for 8+ languages, cutting transcription latency by 35%.

NVIDIA Triton Inference Server FastPitch HiFi GAN Whisper NLLB 200 Batched Inference Beam Search ChromaDB Milvus RAG Vector Search

Research Intern – Generative Models for SAR Imagery

DRDO × JNU Joint Project

Jan 2025 - May 2025 (4 months)

Cut per-epoch training time from 18h to 10h on the PARAM Rudra HPC cluster using PyTorch DDP, gradient accumulation, and mixed precision. Built an unpaired SAR-to-optical translation pipeline with CycleGAN and attention-augmented U-Net generators.

PyTorch DDP Mixed Precision Gradient Accumulation Slurm HPC CycleGAN U Net Synthetic Aperture Radar

Research Intern – Non-Linearity in Deep Learning

Jawaharlal Nehru University

Jun 2024 - Dec 2024 (6 months)

Designed an image augmentation scheme to increase effective non-linearity in CNNs, demonstrating improved generalisation on CIFAR-10. Built multi-GPU training with PyTorch DDP and AMP, achieving a 1.7x speedup.

PyTorch DDP AMP CNN CIFAR 10 Image Augmentation Gradient Flow Profiling