Context

TPQuants needed to transform a rapid prototype built on Replit into a robust, production-grade client-server application for automated Transfer Pricing Audit report generation.

Approach

Led the full rewrite into a React / TypeScript + PostgreSQL application.

Designed and deployed AWS infrastructure from scratch: VPC, EC2 instances, Aurora RDS cluster, and S3 buckets — including network segmentation, IAM policies, and secure connectivity.

Drove ongoing migration, maintenance, and monitoring of cloud resources, establishing repeatable deployment patterns for future engineering team.

Context

Organizations often publish synthetic datasets as a privacy-preserving alternative to real user data. However, generative models can unintentionally learn and reproduce subtle patterns from the original data. These patterns can then be exploited to infer sensitive information, meaning synthetic data is not automatically safe. Evaluating the privacy risk of synthetic tabular data is therefore essential.

Traditional evaluation methods rely on distance-based metrics (e.g., Gower Distance) to simulate attacks between synthetic and real records. While effective for low-dimensional data, these approaches struggle with high-dimensional tabular datasets and fail to capture complex feature interactions.

Approach

To improve privacy evaluation, I investigated whether representation learning can better capture the structure of synthetic records. I built a benchmarking framework that compares: - Anonymeter-style attack simulations using Gower distance. vs - Learned representations obtained from contrastive learning and variational autoencoders (VAE)

The core idea is that learned embeddings should better model feature relationships and improve attack accuracy in high-dimensional settings.

Outcome

Results show that:

- Representation learning approaches can match or exceed the performance of distance-based attacks in high-dimensional datasets.

- Contrastive learning achieves strong detection performance while reducing computation time, demonstrating a scalable alternative for privacy risk assessment.