Google Scholar · GitHub · X

I'm a Privacy Engineer at Google, focused on solving complex privacy challenges at scale. I'm interested in the learning more about AI privacy and security and understanding model awareness + alignment.

Currently thinking about

• Does the model really know it is a model?
• What is the understanding of the model about itself?
• Are the model's answers about "escaping" real or inspired from human text?
• What are better questions to think about?

Experience

• Google: helping solve Ads privacy issues
• Enkrypt AI: built LLM guardrails and internal benchmarks
• Skyflow: 2x intern, secure enclave and transient data
• IBM: intern, secrets vault development

Research

• Privacy-preserving keystroke analysis using fully homomorphic encryption & differential privacy (ICML TPDP 2021)
• SAGE-RT: Synthetic Alignment data Generation for Safety Evaluation and Red Teaming (ArXiv 2024)
• ViT-inception-GAN for image colourisation (ICICC 2021)
• VERA: Validation and Enhancement for Retrieval Augmented systems (ArXiv 2024)
• Robust recovery of adversarial examples (ICML AML 2024)
• Robust Adversarial Training for Detection of Adversarial Samples (ICICV 2022)
• Efficacy of the SAGE-RT Dataset for Model Safety Alignment: A Comparative Study (NeurIPS Pluralistic Alignment Workshop 2024)

Interests

• I enjoy playing chess (≈1900 bullet, ≈1700 blitz and like the Scandinavian Defense.)
• I'm also interested in philosophy, especially Stoicism.
• Currently exploring fundamental questions about AI consciousness and awareness.

Re-inspired by rpandey.tech