Software Engineer I

Data Export Pipeline Optimization

Ruby on Rails

Jira

GitHub

Java

Terraform

Software Engineer II

Savings Automator - cloud cost optimization

Postgres

MySQL

Docker

Kubernetes

Graduate Teaching Assistant

Python

PyTorch

neeraj

A distributed key‑value store that partitions data into shards, replicates each shard with a consensus protocol (Multi‑Paxos) to keep replicas consistent, and supports cross‑shard transactions using two‑phase     commit (2PC). It emphasizes safe reconfiguration (moving shards between groups) and exactly‑once behavior under an asynchronous, crash‑stop model.

Distributed Consensus & Sharded Key‑Value Store

The goal of this work was to evaluate whether LLM could be used to speed up writing the first draft of clinical trial protocols. It combines a large language model (LLM) with “retrieve‑then‑write” steps: it first looks up trusted medical sources, then drafts a protocol that follows standard regulatory structure (FDA and ICH E6). The goal is to reduce manual effort while keeping outputs evidence‑based and easy for clinicians to review.

Conversational AI‑Driven Assistant for Medical Protocol Writing

Domain-specific reasoning project that improves financial‑economics QA by fine‑tuning Llama 3 with a hybrid dataset of textbook‑derived MCQs and bootstrapped rationales. Uses Self‑Taught Reasoner (STaR) and Chain‑of‑Thought to iteratively generate and refine rationales, narrowing the gap between an 8B model and a 70B baseline on multiple benchmarks.

Financial Reasoning Refinement for LLMs

Image‑based static malware detection that converts APK classes.dex bytecode to RGB images, then fine‑tunes EfficientNet‑B4 with a GAP + softmax head. Evaluates 26 pre‑trained CNNs and large‑scale SVM/RF/Stacking; EfficientNet‑B4 achieves state‑of‑the‑art accuracy on test set 1 and strong generalization on a 500‑image real‑world test set 2

EﬃcientNet CNN-based Android Malware Detection

Unsupervised GAN‑based pipeline that learns transferable representations from chest X‑ray and CT images without labels, then classifies diseases using a lightweight stacking classifier. Achieves 94–99.5% accuracies across six tasks with a macro average of 97.6%, reducing annotation needs.

Lung‑GANs: Unsupervised representation learning for lung‑disease classification