Evaluated behavior of GPT-6 Astra and upcoming models in Codex across agentic coding, connectors, CUA, and memory/compaction; calibrated reasoning for token efficiency. Iterated toward a secure, reliable eval environment by stabilizing failure-prone runs, shortening evaluation turnaround, and safeguarding long-horizon agent behavior.

Improved reliability of coding agents through context, tools, delegation, and failure analysis; explored training lower-cost subagents with multi-agent RL.

Worked on early-stage model development behind initial Apple Intelligence launches, including Writing Tools and Reply Suggestions.

sequential recommendation and retrieval; AutoML