PROACTIVE ALIGNMENT. SAFE INTELLIGENCE.

Proposed Research Direction: 

The Question: 

Can we teach models what to learn—and what not to learn—during training itself, rather than correcting them afterwards?

The Concept: 

Current AI development invests massive resources training models on unfiltered data, then applies costly post-hoc fixes for toxicity, bias, and misalignment. This research explores embedding continuous safety guidance directly into LLM training—using curated positive and negative examples to actively steer learning.

Why it Matters: 

Current approaches treat safety as post-hoc correction. Building it into the foundation from the start could be more effective.

Research interest

The Question: 

Can knowledge ever be truly unlearned when it can be reconstructed from other knowledge? If we remove information about making explosives, can that knowledge be recovered from retained physics and chemistry understanding?

Why it Matters: 

Current unlearning approaches test isolated knowledge deletion. But knowledge is interconnected -  unlearning may be insufficient if models can reconstruct harmful information from foundational knowledge that remains.

Research interest / Planned research

The Question: 

How do organizations evaluate third-party agentic AI frameworks before production deployment? What vulnerabilities exist in commercial agent development kits, and what additional safeguards are necessary?

The Concept:

Before deploying agentic AI systems in regulated environments, organizations must evaluate vendor tools for safety and security. This research involves red teaming commercial agent frameworks - testing for prompt injection, jailbreaks, tool misuse, and unintended behaviors.

Why it Matters:

As commercial agent frameworks move from demos to real-world deployment, understanding how they behave in complex environments becomes critical. Proactive evaluation helps organizations unlock their full potential while ensuring safety and trust at scale.