cv

Google DeepMind

Student Researcher, Perception: Humans & Interactions team

• Training ultra-low precision Mixture-of-Experts model for approximating general floating point models.
• Designing efficient quantization schemes for multi-modal generative models with theoretical guarantees.

Texas Instruments

Systems Engineer, High Speed Signal Conditioning team

• Designed an Adaptive Equalizer in RX of Retimer SoCs to reduce Inter Symbol Interference at data rates of 56Gbps.
• Developed a MMSE adaptation algorithm for fast convergence and a steady-state BER of <1e-15.

Texas Instruments

Systems Engineer, Wireless Infrastructure Team

• Developed fixed point models for DDC and DUC chains of 5G Transceiver SoCs and analyzed their SNR and SFDR.
• Identified key contributors of spurs in the DDC chain and devised solutions to maintain SFDR below -100dBc.

PhD

University of Wisconsin-Madison

Electrical and Computer Engineering

MS

University of Wisconsin-Madison

Electrical and Computer Engineering

B.Tech

Indian Institute of Technology, Madras

Electrical Engineering

Implicit Representations via Operator Learning

Proceedings of International Conference on Machine Learning (ICML)

An operator theoretical reformulation of the INR model, able to use CNNs as INRs(which are mainly dominated by MLPs). Allows interpolation between INRs trained on different images while also being able to represent 3D volumes, videos, and 3D MRI scans.

FrameQuant: Flexible Low-Bit Quantization for Transformers

Proceedings of International Conference on Machine Learning (ICML)

A PTQ method for quantizing LLMs and Vision Transformers to effectively two bits with only a small drop in performance. Our method is based on classical frame analysis and provides flexibility to quantize a model to fractional bit widths on average, uniformly across all layers.

Offset Correction in High-Speed Serial Link Receivers

US Patent

Digital Upconverter for Radio Frequency Sampling Transmitter

US Patent

Lane Adaptation in High-Speed Serial Links

US Patent

Data Augmentation for Image Classification using Latent Diffusion Models

• Generated synthetic images using Stable Diffusion for CIFAR-10 dataset and trained a ResNet18 model.
• Fine-tuned the Stable Diffusion using Textual Inversion, Dream-Booth to improve the accuracy to ∼82%

Performance of Multi-Image Comparison Queries

• Proved that the Cross Cluster Edge Density for queries with 3 to 8 images per query is better than the classical pair comparison queries under Conditional Block Model.
• Implemented a Python program to generate mathematical proof and derive the expressions for cross cluster edge density given the number of images per query.

Efficiency Model for Buck Converter

• Developed an Empirical model for the total power loss in a Buck converter with maximum error <3%.
• Derived an analytical expression for the Optimum sizes of switches and the frequency of operation of the Buck Converter.

Robust Image Classification

• Processed noisy imaged with DCT, DWT and improved the classification accuracy by 15% on the CIFAR-10 dataset.
• Implemented a Bayesian uncertainty-based two-layer defense to defend a classifier against Adversarial attacks such as FGSM, BIM, and CW L2.

Systematic Encoding and Decoding of BCH Codes

• Designed a systematic encoder and decoder for a binary BCH code with error and erasure correcting capabilities.
• Implemented the simplified Berlekamp-Massey algorithm for binary codes to determine the error location polynomial.