Looking for a job

Ankur Kumar User

@ankurkumaruser

Message

Data Scientist at Oriserve who reduced Whisper ASR hallucinations by 90% and deployed production GenAI assistants.

India

Message

At Oriserve, I fine-tuned speech and ASR models for Indian multilingual telephony, including Whisper, Qwen-ASR, and NVIDIA NeMo. I applied a calm-whisper technique that reduced Whisper ASR hallucinations by 90% and used weighted cross entropy to reduce WER by 15%.

I also built production GenAI solutions for EF Education First and TVS Motor using LLMs, RAG, FastAPI, and Pinecone. In research at the Machine Learning and Geocomputing Lab, IIT ISM Dhanbad, I trained diffusion models on OpenFWI seismic data and achieved 86% accuracy in seismic FWI.

Experience

Work history, roles, and key accomplishments

Data Scientist

Oriserve

Aug 2025 - Jul 2026 (11 months)

Fine-tuned speech and ASR models for Indian acoustics, reducing hallucination by 90% and WER by 15%. Built production GenAI solutions for clients using LLMs, RAG, and FastAPI, and deployed models with vLLM.

Machine Learning Deep Learning NLP Speech Recognition Whisper Qwen ASR NVIDIA NeMo PEFT LoRAs QLoRA GRPO Knowledge Distillation VLLM RAG fastAPI Pinecone Amazon Bedrock Docker Jenkins BERT NER AWS

Research Scientist

Machine Learning and Geocomputing Lab, IIT ISM Dhanbad

Aug 2024 - Feb 2025 (6 months)

Pretrained a Denoising Diffusion Probabilistic Model for seismic velocity synthesis and achieved 86% accuracy in seismic FWI with a Physics Informed Attention UNet.

OpenCV PyTorch Diffusers Deep Learning Computer Vision DDPM FWI

Education

Degrees, certifications, and relevant coursework

Indian Institute of Technology (Indian School of Mines)

M.Sc.Tech., Applied Geophysics

2022 - 2025

Grade: 7.71

Pursued a Master's degree in Applied Geophysics with relevant coursework in Deep Learning, Digital Signal Processing, and Remote Sensing. Completed a thesis on Seismic FWI using Deep Learning with Diffusion Models and Spatial Attention U-Net.