Applied Researcher · Prime Intellect
I’m currently an Applied Researcher at Prime Intellect, where I work on post-training, synthetic data, and benchmarks for building robust long-horizon agents. My research focuses on reinforcement learning, instruction tuning, verifiers, and evaluation, with a particular emphasis on reward hacking and verifier failure modes in large-scale RL environments and post-training runs. Previously at Microsoft, I led applied research across Microsoft 365 Copilot and Frontier Tuning, working on post-training, reasoning, personalization, hallucination detection, and domain-adaptive agents.
Current focus
My research focuses on self-improving agents: how they can generate useful experience, learn continuously from interaction, and become more capable over time. I’m particularly interested in the learning signals that enable this—rewards, verifiers, credit assignment, and trajectory-level evaluation—and in making them robust to reward hacking and other failure modes.
I work across the post-training loop—supervised fine-tuning, on-policy data, reinforcement learning, synthetic task generation, environments, and evaluation—to improve agents that plan, use tools, recover from mistakes, and make progress over many steps. I’m especially interested in trajectory-level credit assignment and whether learned behaviors generalize across tasks, environments, and tool interfaces.
Automated rewards make large-scale RL possible, but they also create new failure modes. I build and stress-test verifiers, compare outcome and trajectory-level signals, and look for proxy optimization, reward hacking, and brittle evaluation before those signals are used to train a stronger policy.
Selected work
Selected work across Prime Intellect and Microsoft, spanning agentic RL, evaluation, reasoning, personalization, and production ML.
Built an experimental workflow for testing whether an automated reward remains trustworthy under optimization pressure. The approach uses controlled probes, defense ablations, hardened re-scoring, and trajectory inspection to surface verifier brittleness and reward hacking before scaling an RL run. I open-sourced the reference implementation for SWE-style agent tasks and presented the work at Ray Summit 2026.
Another focus of my work is scalable synthetic task generation for training and evaluating long-horizon agents. I build pipelines for generating realistic, multi-step workflows in simulated tool environments, with structured world states and programmatically verifiable outcomes. I’m particularly interested in generating tasks that span a wide range of difficulty and expose failures in planning, tool use, and long-horizon execution.
Code · TalkLed applied research on post-training and evaluation for Microsoft 365 Copilot, including on-policy distillation and training for tool-using reasoning systems. The work focused on improving response quality, citation accuracy, tool use, and reasoning on complex enterprise tasks while making the resulting behaviors easier to evaluate and iterate on.
Microsoft 365 CopilotLed post-training for a domain-adaptive reasoning agent in Copilot Tuning that formulates planning tasks as optimization problems. I worked across training, synthetic data, and evaluation for multi-turn tasks where the model has to reason over constraints rather than produce a single short answer.
Copilot Tuning OverviewDeveloped an LLM routing framework for hallucination detection in Microsoft 365 Copilot that reduced evaluation cost by 60% and latency by 30% without degrading accuracy. The work combined model-quality measurement with the systems constraints required to run evaluators at production scale.
Helped ship Copilot memory and built an LLM-based evaluator for measuring personalization across conversations. This work focused on whether a model uses remembered context in ways that are relevant, consistent, and actually helpful to the user.
Copilot Memory AnnouncementEarlier in my Microsoft work, I shipped transformer-based user embeddings that improved file-recommendation recall by 23%, and led LoRA-based smart-reply work that increased click-through rate by 3% while model compression reduced production latency by 35%. These projects shaped my interest in closing the loop between research, evaluation, and real user-facing systems.
Related PublicationResearch
Research spanning agent evaluation, model safety, grounded generation, and sequential decision-making.
A framework for stress-testing agent verifiers under controlled changes to tasks, tools, rubrics, and observations. TRACE measures brittleness, transfer, proxy divergence, reward-hacking risk, and process overfitting before verifier scores are used for training or deployment.
Introduces a permissive information-flow framework for controlling sensitive information in language-model outputs while preserving useful model behavior.
PaperProposes a decoding method that incorporates external knowledge during generation, enabling language models to use grounded evidence without requiring the knowledge to be baked into model parameters.
PaperStudies how semantic representations of emotion labels can improve prediction of emotional reactions in text, particularly when labels have meaningful structure that standard classifiers ignore.
PaperSpeaking
Selected presentations on agent evaluation, post-training, and reinforcement learning.
Women Who Code Summit, New York — June 3, 2026
Design patterns for using agentic evaluators to scale judgment beyond static LLM-as-a-judge pipelines, with a focus on reliability in real enterprise systems.
Session DetailsAgentic AI Summit, UC Berkeley RDI — August 2, 2026
Work on detecting brittle verifier behavior when long-horizon agents are trained against automated rewards.
Ray Summit, San Francisco — August 25, 2026
A systems-focused discussion of reinforcement-learning training for large language models and the role of high-throughput inference in the post-training stack.
Session DetailsRay Summit, San Francisco — August 26, 2026
A practical workflow for trying to break an automated reward before spending a large RL run optimizing it: controlled probes, defense ablations, hardened re-scoring, and trajectory-level audits for SWE agents.
Session Details · CodeService
Industry track reviewer for major NLP conferences, including: