I specialize in privacy-enhancing technologies and the applied cryptography that makes them practically deployable, focusing on differential privacy, threshold protocols, and verifiable computation. My work bridges the gap between theoretical guarantees and production realities, whether that means engineering secure AI guardrails or exposing hidden flaws in standard privacy assumptions.

In my independent research on differentially private learning for imbalanced clinical data, I found that standard approaches often fail quietly. Uniform gradient clipping is not class-neutral: it systematically suppresses minority classes, with median per-class gradient norms diverging by a factor of 7.75. Furthermore, common remedies like SMOTE silently break record-level privacy guarantees by clustering synthetic and real patients into correlated groups. To fix this, I introduced per-class adaptive bounds that improved AUC without diverging across thousands of runs, demonstrating that the assumed privacy-utility tradeoff is not always a strict zero-sum game. The working paper is on the research page, with code at github.com/sha256rma/dp-diabetes-prediction.

I am equally focused on the architectural deployment of secure systems. During my graduate studies at the City College of New York, I published six peer-reviewed papers (earning a Best Presentation Award at IEEE UEMCON 2021) on patient-controlled health record sharing. I acted as the sole implementer for a system combining threshold proxy re-encryption, Ethereum access-control contracts, and multi-node private IPFS. Beyond patching vulnerabilities via Slither pre-publication, I navigated the deployment through Ethereum's proof-of-work to proof-of-stake transition, re-deriving timing logic to achieve a nearly 80% latency reduction. Rather than hiding system limitations, I explicitly documented the friction points of real-world cryptography, such as the forward-only nature of revocation and the inherent metadata leakage of relationship graphs on-chain.

In the industry, I recently owned a production guardrail suite of seven detection models as an AI Security Engineer at Tumeryk. I cut mean Safety Check latency by 58% with zero accuracy loss by migrating serving to vLLM on GCP, and drastically reduced bias false positives from 25% down to 2%. I also built a Garak-based red-teaming harness, exposing severe reproducibility issues in LLM-as-judge evaluations where attack-pass rates varied from 6% to 91% based solely on judge configuration, and authored formal public comments proposing MITRE ATLAS integration for NIST IR 8596.

My engineering is grounded in deep academic roots, including studying PhD-level Modern Cryptography and Zero-Knowledge Proofs. Today, I actively share that foundation by teaching the cryptographic primitives of Bitcoin, from one-way functions to threshold signing with FROST, as a Bitshala Educator Fellow.

Education

Research Areas