Applications Software Engineer at Oracle
Bengaluru, Karnataka, India
Selected work
I enjoy the unglamorous middle: logs, test cases, and the stack trace everyone has been avoiding. These are the projects I’m happiest to talk about.
Tools that earn their keep on a Monday morning.
Production intelligence at Oracle
What was messy. Following a production issue meant jumping between logs, database context, and distributed code paths.
What I made. An orchestrator-worker multi-agent AI system that pulls together OCI log analysis, SQL workflows, and code intelligence. It starts with evidence and keeps the guesswork out of the loop.
What changed Investigation time dropped by 70%.
Quality engineering at Oracle
What was messy. Good regression coverage took too much manual effort, especially for edge cases and unhappy paths.
What I made. An AI-assisted platform for finding scenarios, drafting test steps, and running them.
What changed Manual QA effort dropped by 80%
Research project
What was messy. Old COBOL code often knows more about the business than its documentation does.
What I made. A Python framework for finding and explaining business rules, with fine-tuned Gemma, Llama 2, and Mistral models.
What changed Extraction accuracy improved by 32% over baseline models.
View the repository for COBOL RuleForgeExperience
Research and education
Developed an LLM-based extraction framework with an F1 score of 0.73 versus COBREX’s 0.59, and 80%+ user preference in comprehension studies.
Built a COBOL85 grammar and AST-based translator with 98.35% computational accuracy across 103 CodeNet COBOL programs.
Compared DCGAN, SRGAN, and CGAN models using CIFAR-10, MNIST, GAN loss, and Fréchet Inception Distance.
Evaluated machine-learning methods for weather forecasting, achieving a 95.70% regression score with a Decision Tree Regressor.
Skills
