Staff Research Scientist at Google DeepMind, working on post-training and agents. Currently leading post-training for Legal AI — including data for SFT and RL, recursive self-improvement, human/SME annotation, and learning from production and post-deployment signal. Previously Director of Research at Contextual AI, managing a team of researchers working on LLMs and agents, and an Applied Scientist at Amazon, where I helped found the Alexa LLM post-training team (which became Amazon AGI). PhD from CMU LTI [thesis] (2022) and BSc from UBC (2018).
At Contextual AI: (1) Reflective Context Learning accepted at COLM 2026 (2) Grounded Language Model reached #1 on the FACTS leaderboard (VentureBeat) (3) LMUnit open-sourced and achieved #1 on RewardBench2 (4) AgentLens launched as a multi-agent evaluation system (5) Post-training recipe featured as a Google Cloud case study
I most enjoy product- and impact-driven research — work that enables next-generation user experiences through strong problem formulation. Key areas include:
- Post-training/Learning: RL and alignment [apo '24, nl feedback '23], learning from post-deployment signal [life-long learning '25], continuous learning/context optimization [rcl '26], RL for agentic search, factual grounding [glm '24, #1 on FACTS leaderboard]
- Data: Data factories as the main lever for model behavior [age of research '26], synthetic data [model collapse '24], human/SME annotation, model-in-the-loop data collection [lad '22], instruction data [instructdial '22, cesar '23], and specialization/custom finetuning
- Agents/Compound AI Systems: Agents and agentic workflows [workflows → agents '26], multi-agent systems [league of legends '25], multi-step pipelines [structured fusion '19], retrieval-augmented generation [rag 2.0 '24], structured representations [schema-guided '21], end-to-end optimization
- Evaluation: Agentic evaluation [agentlens '25], verifiers and natural language unit tests [lmunit '24, verification vs. generation '25], user simulation and user-agent collaboration [ugst '25, accord '26], automatic metrics [usr '20, fed '20], realistic user experience evaluation [nsf report '22, interactive eval track '22, dialport]
During