Neel Vora

I'm a computer science graduate pursuing Master's degree. I bring over three years of experince in field of applied machine learning and software engineering. Currently I am machine learning intern at Lawerence Berkeley National Laboratory in Berkeley, where I work on optimizing quantum qubit readout through signal processing and noise reduction techniques.

I am currently working on my thesis under the guidance of Dr. VP Nguyen specializing in the field of machine learning on edge devices.

Email  /  Resume /  Github /  LinkedIn  /  Scholar

Highlighted Works

I'm interested in computer vision, machine learning, image and signal processing processing, edge computing. Most of my work is about development and deployment of end-to-end machine learning pipeline either on edge device or cloud servers.

This project swiftly classified quantum qubit states using ML algorithms. Raw voltage signals from a cryogenically stored quantum chip were processed, and noise removal and dimension reduction were performed. The ML algorithms, along with a data-driven digital local oscillator, enabled high-fidelity classification—all implemented on FPGAs, completing the process in just 2µs.

This project addresses limitations in camera-based tracking systems by implementing a multi-modal approach, combining both acoustic and vision tracking. Leveraging student-teacher learning, the acoustic model was trained from the pre-trained vision model. The integration of a cross-attention mechanism maximized the advantages of both modalities. The project demonstrates the efficacy of this approach in UAV detection and tracking. Implemented on the ARM Cortex A72, the system ensures lightweight and real-time inference for practical applications.

This project focuses on extracting valuable information from physiological signals and compressing their size. Utilizing a variational auto-encoder, traditionally employed for images, the project compresses physiological signals. To showcase its practical application, the approach was implemented on a real-life seizure patient. The compressed signals were then utilized for seizure detection as a downstream task, validating the effectiveness of the compression method. The entire system was implemented on ARM Cortex A72 and Cortex A57, highlighting real-time capabilities.

Experiences

At LBNL, I optimized superconducting quantum qubit readout, reducing time from 4μs to 1.5µs using signal processing and noise reduction. Developed a Pytorch LSTM achieving 98% fidelity for discriminating 8-qubit states. Applied agile practices with TensorFlow, Kubernetes, and Grafana. Translated the model to HDL and implemented it on an FPGA with model inference time of just 24ns.

At WSSLab, I achieved a 1:293 compression ratio for EEG signals using VAE. Implemented a real-time data pipeline with ARM Cortex and Nvidia Jetson. Developed a scalable patient's seizure monitoring and detection system deployed GCP using Node.js and machine learning. Created interactive interfaces with React and established MongoDB backend. Validated the end-to-end system for EEG data through quantitative metrics and real-world applications

Domain of Object Tracking
UTA, I enhanced UAV detection by 26% using cross-modal self-supervised learning, surpassing state-of-the-art models in non-line-of-sight scenarios. Created a unique CRNN-based acoustic model and employed YOLO as a teacher model for UAV detection in low-light (1.75 Lux) and Blockage conditions.
Domain of Genome
Implemented GANs to reconstruct the g-carbon distance matrix and utilized ADMM for 3D protein structure conversion, enabling more accurate and efficient representation of molecular structures for advanced biomedical applications.
Domain of NLP
Improved LLM (llama with 13 billion parameters) inference throughput with a paged attention technique. Established a Kubernetes-based pipeline, reducing workflow latency by 15% through efficient data management. Collaborated with a cross-functional team to integrate the app with existing systems, boosting system efficiency by 10%.

Research

The paper presents DroneChase, a mobile drone tracking system using a camera and microphone array that achieves robust performance under obscured conditions by fusing acoustic and visual modalities in a self-supervised, cross-modality framework

This paper proposes a variational autoencoder architecture that compresses multi-channel physiological signals by over 293x while retaining 91% seizure detection accuracy. Validated on real EEG data from epilepsy patients, it demonstrates clinical utility and enables substantial power savings of up to 26.8% on edge devices.

EarSD applies lightweight ML models including SVM, KNN, and Random Forests to detect epileptic seizures. The models are trained on processed EEG, EMG, and EOG signals acquired from behind the ears of 33 epilepsy patients. They achieve up to 95.3% seizure detection accuracy, on par with video-EEG.

Thanks Jon Barron, for this website template.