Selected work.
From training models to building data platforms and evaluating how agents interact.
INDEPENDENT RESEARCH / 2026 — ONGOING
Parley Labs
Built environments to study and evaluate multi-agent negotiation, collecting interaction data and decision traces across marketplace and Catan experiments.
Marketplace
- Studied how agents buy and sell goods, negotiate prices, and reach agreements.
Catan trading agents
- Studied how agents plan their moves and negotiate resource trades as part of a strategy to win Catan.
CATAN® is a trademark of CATAN GmbH. The original game was designed by Klaus Teuber. This independent research is not affiliated with or endorsed by CATAN GmbH or CATAN Studio.
Independent research in collaboration with Milind Shah.
EMPLOYMENT / SELECTED CONTRIBUTIONS
AI in practice.
Midstream
Post-training and evaluating extraction models.
- Lead document extraction and contract ingestion for 8+ health systems.
- Work on post-training models for document extraction.
- Turn execution traces and human feedback into evaluation datasets and better baselines.
Pipelines
Building the AI R&D cycle.
- Built data collection, generation, and evaluation into one platform as the first engineer.
- Connected workflows with programmatic checks, LLM judges, and human review.
Meta
Training models. Shipping their impact.
- Improved follow prediction accuracy and stability by 21% through training data and scoring changes.
- Built voice and NLP integrations for Meta Ray-Ban and agentic workflow automation.
More experience, education, and skills are in my résumé ↗.