Research

Publications

2026

Sandwich-Residuals: Parameter-Efficient Test-time Adaptation of World Models

Krishnam Soni, Aditya Sehgal, Vedant Dave, Elmar Rueckert

Adapts a frozen world model via small predictor residuals, with 97–99% fewer adapted parameters.

Submitted to ICRA 2027

Paper ↗Website ↗

ALRA: Action-Latent Residual Adaptation for World Models

Krishnam Soni, Aditya Sehgal, Vedant Dave, Elmar Rueckert

Corrects dynamics shifts with a 210-parameter affine correction to a frozen world model's actions.

NeurIPS 2026 Workshops: PTA, TTCL, CL4FMAgents

Paper ↗Website ↗Code ↗

PULSE: Surprise-Based Online Unsupervised Keyframe Detection

Aditya Sehgal, Krishnam Soni, Elmar Rueckert, Vedant Dave

Online, unsupervised keyframe detection from visual and proprioceptive predictive surprise.

NeurIPS 2026 Workshops: RoboPAD, PTA

Paper ↗Website ↗

2024

Retinal OCT scan next to its colour-coded layer segmentation mask

Segmenting Optical Coherence Tomography (OCT) Scans Using Multi-Scale Modified UNet Architecture

Krishnam Soni, Karri Karthik, Manjunatha Mahadevappa

Retinal-layer segmentation in OCT with a multi-scale UNet, surpassing inter-observer agreement.

IEEE INDICON 2024

Paper ↗

Academic service

Reviewer: NeurIPS 2026 Workshops (WM PAI, TTCL)

Other projects

Image Denoising with UNet

UNet denoising on CBSD68 (σ=35, augmented to 204 images). A W&B sweep compared Batch, Instance, Group and Layer normalization with MAE and MSE losses. The best model (MAE + LayerNorm) reached a validation PSNR of 27.50 and SSIM of 0.784.

Python · TensorFlow/Keras · Weights & Biases

GitHub ↗

ViziMed

A Flask medical management system for secure patient data handling and diagnosis assistance, with user authentication, patient logs and medical scan uploads. The frontend is written from scratch in HTML, CSS and JavaScript, without any frontend libraries.

Python · Flask · SQLAlchemy · HTML/CSS/JS

GitHub ↗Live demo ↗