Esha Pahwa

I am currently working as an AI Software Engineer at Corvic AI, where I build and deploy production-grade LLM and retrieval-based agentic systems for enterprise applications, with a focus on grounded generation, evaluation, and system reliability. I am an MS in Machine Learning (MSML) graduate from Carnegie Mellon University. Previously, I was a Member of Technical Staff at Adobe, and a Research Associate at Google Research India in the Shopping Ads team, where I worked on large-scale product retrieval and ranking using JAX, mentored by Gaurav Srivastava and Prateek Jain.

My research background spans computer vision, multimodal learning, and representation learning. I conducted research at Adobe Research with Balaji Krishnamurthy on lookalike modeling and segmentation, and worked on image restoration and super-resolution with Prof. Pratik Narang (IISc) and the VCG Group at Harvard University under Prof. Hanspeter Pfister, mentored by Salma Abdel Magid. I have publications at venues including NeurIPS workshops, AAAI, and ICCV.

Email  /  CV  /  Google Scholar  /  Linkedin  /  Github

News

Research

I'm interested in computer vision, natural language processing, and machine learning. Much of my research is based on the applications of these topics in real-world scenarios.

SITCOM augments pretrained Vision-Language-Action models with model-based rollouts and reward-guided planning at inference time, enabling robust long-horizon control. This approach improves task success from 48% to 72% in simulated robotic environments.

We generate the LowVis-AFO dataset, containing 3647 paired dark-hazy and clear images. We also introduce a new lightweight deep learning model called Low-Visibility Restoration Network (LVRNet)

We propose a novel Sparse Weighted Temporal Attention (SWTA) module to utilize sparsely sampled video frames for obtaining global weighted temporal attention. The SWTA network can be used as a plug-in module to the existing deep CNN architectures, for optimizing them to learn temporal information by eliminating the need for a separate temporal stream.

This is the extended version of the study conducted for the proposed SWTA network. It contains a detailed survey of the existing approaches with an elaborate explanation of the SWTA network and its components.

We introduce a novel architecture Toggle-Fusion Network (TFNet) that effectively utilises a multimodal dataset, RGBT-CC, containing pairs of thermal and RGB images.

We propose a novel architecture of a modified HRNet which includes added skip connections along with convolutional block attention modules (CBAM). We evaluate our model on two publicly available datasets, PEIR Gross and IU X-Ray.

Mini Projects

This project was completed in collaboration with CSIR-CEERI. Utilized OpenPose API to extract precise joint coordinates from video datasets. An LSTM-based model trained on raw joint coordinate data, optimizing the training process for accurate yoga pose prediction due to the unprocessed nature of the input data.

Participated in Pytorch Summer Hackathon 2020 with my batchmate Achleshwar Luthra. We used a pre-trained AlexNet model for classifying if the patient had the aforementioned disease or not through the image of their eye. The model was later deployed using Flask framework.

Borrowed from Jon Barron website's source code. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead. Also, consider using Leonid Keselman's Jekyll fork of this page.