Looking for a job

Naman Mishra

@namanmishra5

Machine learning research engineer at CVIT building efficient world models for hazardous driving scenarios and Vision-Language Models.

India

Message

At CVIT, International Institute of Information Technology Hyderabad, I'm developing and benchmarking efficient world models for hazardous and edge-case driving rollouts, using 3D-geometry conditioning and feedback from latent world models. I'm also developing Vision-Language Models for autonomous driving and modeling pedestrian behavior.

At Eagle Eye Networks, I instruction fine-tuned Vision-Language Models to reduce object hallucinations in weapons detection and developed models for person activity recognition and dense captioning. Earlier, at Uncanny Vision, I designed and evaluated Scene Text Recognition models for License Plate Recognition and created synthetic training datasets.

Experience

Work history, roles, and key accomplishments

Current

Research Fellow

Current

CVIT, International Institute of Information Technology Hyde

Mar 2025 - Present (1 year 7 months)

Developing and benchmarking efficient world models for hazardous and edge-case driving rollouts through 3D-geometry conditioning and feedback from latent world models. Developing Vision-Language Models for autonomous driving and modeling pedestrian behavior in structured and unstructured traffic environments.

World Models 3D Geometry Vision Language Models Autonomous Driving Pedestrian Behavior Modeling

Senior Machine Learning Research Engineer

Eagle Eye Networks

Sep 2021 - Jan 2025 (3 years 4 months)

Instruction fine-tuned Vision-Language Models (VLMs) to reduce object hallucinations for weapons detection, using diverse negative samples, high-resolution vision encoders, auxiliary dense captioning, and a custom QFormer projector. Developed an end-to-end VLM for person activity recognition, using dual ViT-Base encoders and lightweight LLMs to overcome overfitting, achieving robust performance.

Vision Language Models Instruction Fine Tuning Object Detection Dense Captioning QFormer ViT LLM Panoptic Segmentation Scene Graphs Scene Text Recognition Synthetic Data Generation Data Curation

Senior Machine Learning Research Engineer

Uncanny Vision

Sep 2019 - Sep 2021 (2 years)

Designed, implemented, trained, and evaluated a Scene Text Recognition model for License Plate Recognition (LPR) for applications in vehicle analytics. Introduced an entropy-regularized CTC function for more robust STR performance and pruned the feature extraction backbone for reduced latency.

Scene Text Recognition License Plate Recognition CTC Model Pruning Synthetic Data Generation YOLOv3 Mask RCNNs Grad CAM

Machine Learning Research Engineer

Uncanny Vision

May 2018 - Sep 2019 (1 year 4 months)

Refactored the CRNN repository from Lua to PyTorch and the YOLOv3 object detection repository from Darknet to PyTorch, obtaining similar accuracy and throughput. Implemented scripts for affine and non-affine transformations of images and labels, and orchestrated dataset collection and curation.

PyTorch CRNN YOLOv3 Darknet Image Transformation Data Curation Team Supervision