Projects that blend research and engineering.

5 attacks and 3 defenses, benchmarked head-to-head on GT's HPC cluster.

Developed a solo agentic benchmarking platform (Python, LangChain, Slurm on the Georgia Tech HPC cluster) implementing 5 attacks and 3 defenses drawn from 5 papers. The runs revealed that 2 attacks defeated all defenses while 1 defense failed against basic prompt injection — grounding ongoing research into dynamic ML security defenses.

ML Security Research

Recovered 25% robust accuracy on perturbed images with a PGD pipeline.

Built a PyTorch adversarial fine-tuning pipeline for ResNet50 that recovered accuracy on perturbed images by 25%. Streamed ImageNet-1k through a GPU-accelerated adversarial-robustness pipeline performing 20-step PGD attacks on 5K+ samples.

ML Research · Python · PyTorch

Reconstructing authenticating feature vectors to test a 105-bit iris key scheme.

Designing an ML attack in PyTorch against a 105-bit iris-based key-derivation scheme, exploiting subsample Hamming-distance correlations to reconstruct authenticating feature vectors. Using a Stable Diffusion model evaluated on the IITD and ND iris datasets to test whether the scheme's effective entropy falls below its claimed bound.

ML Security Research · Georgia Tech

Raised measured groundedness to 95% across 1,200 chats over dense SEC filings.

At FINRA, built an automated evaluation pipeline on AWS Bedrock and S3 that scored precision, recall, and used LLM-as-a-judge at the chunk and document level to guide retrieval and prompt improvements — raising measured groundedness to 95% across 1,200 internal chats. Also developed an MCP extension that models task-completion time from live conversation context, quantifying ~8 hrs/wk in productivity gains across 100 users.

Graduate AI Engineering Intern · FINRA

Cut a multi-hour manual DNS process down to sub-5-second automated responses.

At VeriSign, architected and implemented a Spring Boot RESTful API to automate domain-name data retrieval, enabling engineering teams to fine-tune DNS ML systems and generate performance-analysis reports for latency checks. Integrated Spring Security with a custom login filter and token-based authentication, combined with server-side caching and dynamic rate-limiting.

Product Engineering Intern · VeriSign