Differential Privacy

Synthetic Data

Large Language Models

Built an event-driven ETL pipeline — S3 ➔ SNS ➔ SQS ➔ Lambda ➔ Snowflake — ingesting ≈ 17 GB/month of compliance files with exactly-once semantics via idempotent S3 object-tag checks.

Designed and automated regional fail-over for the entire stack (S3 ➔ SNS, Fargate BFF API, EventBridge + Batch, OpenSearch), authoring 4 runbooks and 2 Python scripts with unit/e2e tests; cut recovery time from 40 min to < 5 min.

Led a feature-flag A/B experiment (Segment) on two UI variants across 100 M users; Snowflake analysis showed ~125 % click-through for Variant B, guiding product-pilot rollout.

Automated operational alerts with CloudWatch metric filters and SNS ➔ PagerDuty hooks that trigger on any HTTP 5xx in a 5-minute window and for OpenSearch CPU/Mem thresholds.

Hardened security & compliance: provisioned least-privilege IAM roles and KMS-encrypted S3 buckets / SQS queues; passed 2024 audit with zero critical findings.

Mentored six engineers on AWS best practices and code reviews; 6 out of 6 have acquired the AWS SAA certificate.

Developed a proprietary tool using machine learning to predict B2B credit card acceptance and support bank relationship managers to develop stickier relationships with small businesses.

This innovation contributed directly to over ~$3+ billion in quarterly expenditures.

Led a machine learning project to forecast delinquency risks in business accounts and devise strategies to minimize overcharges.

Oversaw workflow, including model development, monitoring, and performance evaluation.

We give new differentially private learning algorithms for halfspaces and decision lists in the online and PAC models. In the online model, we give a private analog of the classic Winnow algorithm for learning large-margin halfspaces. As an application, we describe how to privately learn decision lists in the online model, qualitatively matching state-of-the art non-private guarantees. We also give a learner for decision lists in the PAC model, where we give a computationally efficient algorithm with similar sample complexity as the best non-private algorithms.

Algorithm development and theoretical analysis of differentially private (DP) learners in the batch (PAC) and online settings for polynomial learnable concept classes (i.e. generalized decision lists, sparse disjunctions, halfspaces, etc.)

PAC setting: Developed a novel DP and computationally efficient algorithm for generalized decision lists with optimal sample complexity

Online setting (oblivious adversaries): Identified adaptation of DP algorithm for expert prediction that learns k-width disjunctions.

Online setting (oblivious adversaries): Developed a variation of the Winnow algorithm that learns k-width disjunctions with high probability regret bound O(klog(d)). Identified two other variants with optimal regret.

Advisor: Mark Bun

Collaborators: None