Resume

• Spearheaded end-to-end delivery of core data protection features for Google Cloud BackupDR, owning the lifecycle from design through production deployment.
• Engineered query performance analysis to profile data scenarios; leveraged query and DB-level optimizations to drive p99 latency for complex queries on millions of records down to sub-500ms.
• Extended system capabilities into Infrastructure-as-Code, implementing declarative Terraform solutions for large-scale restore operations.
• Implemented backend services in Golang and state-managed Angular frontends; hardened system security through critical vulnerability remediations.

Worked with Prof. Praveen Paruchuri on problems spanning machine learning, cybersecurity, and mathematical optimization - specifically Moving Target Defense using factored MDPs, multi-armed bandits, and game-theoretic models. Published at AAMAS 2022, AAMAS 2024, and COMSNETS 2025.

• Built a resource integrity validation, reporting, and bug-filing system in Golang from scratch by analyzing operational workflows across multiple Spanner databases via gRPC.
• Automated reporting resolved issues impacting over 10% of GCP management server instances, significantly reducing SWE debugging time on recurring inconsistencies.
• Prototyped an internal AI chatbot using large language models and Duet AI to enable natural-language querying of backup and disaster recovery job data.

Worked with Prof. Vineeth Balasubramanian on open-world object detection using one-stage models (YOLO, SSD, RetinaNet) - a problem previously only tackled with two-stage models. Implemented contrastive clustering and energy-based classification to detect and label unknown objects, enabling incremental learning of new classes. Built an end-to-end training pipeline.

• Benchmarked neural and statistical models for on-device spellchecking in Google Workspace editors under resource constraints in encrypted documents.
• Enhanced on-device spellcheck precision by 15.1% and recall by 14.2% across five languages using real-time context passing and NLP-based re-ranking in Java; unblocked internal launch.
• Detected and fixed high-priority bugs; executed rigorous unit testing with JUnit.

Implemented novel bandit-based learning algorithms to generate adaptive defender strategies for Moving Target Defense in web applications and networks. Evaluated on real-world National Vulnerability Database (NVD) datasets against SOTA approaches - outperformed baselines while operating under strictly fewer assumptions about the attacker model.

Built a multi-agent Q-learning framework for Demand Response in smart grids, coordinating household appliances under peak power constraints. Achieved 44.7% energy cost reduction over unoptimized baselines while satisfying hard grid-level power limits and minimizing user dissatisfaction.

Fine-tuned domain-specific transformer models (SapBERT, RoBERTa, PubMedBERT) with a classifier layer for NER of disease mentions in clinical case documents. Implemented entity linking with cosine and euclidean distance metrics over learned embeddings to resolve mentions to canonical medical concepts.

Built a full-text search engine over the complete English Wikipedia dump (90GB). Constructed an inverted index compressed to 1/4 of the original data size, supporting field-based search (title, body, categories, infobox) with TF-IDF ranking and sub-second query latency.