I build multi-agent AI systems and train the frontier models underneath them.

I co-founded actAVA and own its AI organization. actAVA is a workflow-to-model learning platform for healthcare: we give enterprises agentic sovereignty — the ability to build, test, and own their AI agents and models.

I lead post-training of Cura 1T, our trillion-parameter healthcare model, trained through a recursive self-improvement loop in which a training agent finds the model's capability gaps and closes them with SFT, RL, and self-distillation. Cura ranks at or near the top of frontier baselines on five of six healthcare evaluation panels while holding its out-of-domain reasoning.

I also lead χ-Bench, our benchmark for long-horizon, policy-rich clinical workflows, where the best frontier agent resolves only 28% of tasks. On the platform side I work on KORA, a model-independent harness for building, testing, and continually improving agents, and CHRYSO, which enforces AI governance against 85+ regulatory controls.

Previously at Salesforce AI Research, I led the post-training team. We shipped xLAM, a family of state-of-the-art function-calling models, the synthetic data pipelines behind APIGen, APIGen-MT, and AgentOhana, and Retroformer, a generative critic model for self-reflection.

I also built agent systems there, including SlackAgents, AgentLite, CodeGenie, SWE Agents, and AIOps for cloud incident causation analysis.

I did my Ph.D. and M.S. in Machine Learning at CMU, focusing on model interpretability, where I was advised by Dr. Kun Zhang.