founder & ai scientist — new york
I build AI systems that learn continuously — and act optimally.
Co-founder & CTO of Greenshoe. Previously at AWS AI Labs, where I trained Amazon Titan models and pioneered the memory and RAG systems behind Amazon Bedrock and Amazon Bedrock AgentCore. PhD in computer science from the University of Pennsylvania.
Greenshoe is the AI-native platform for SEC disclosures: it drafts, benchmarks, and reviews 10-Ks, 10-Qs, and IPO filings — weeks of manual work compressed into hours — and continuously monitors market and regulatory signals so public companies and their law firms act in real time, not at the deadline. It's my research thesis in production: a system that learns continuously and acts optimally.
$3M seed led by AIX Ventures. We're hiring — talk to me.
Most AI systems stop improving the day they ship: experience is thrown away, reasoning is recomputed from scratch, data goes unread. My research treats each as an optimization problem — learning from experience optimizes the model, planning optimizes the use of compute, and retrieval optimizes the knowledge every decision draws on.
Turning the knowledge, experience, and reflections an agent accumulates in memory into training signal — reinforcement learning that makes models and agents improve with use, from RLHF for grounded generation to agents that learn from their own past interactions.
Amazon Titan · Bedrock AgentCore Memory
Scaling reasoning at test time and deciding where compute goes — across steps, tools, and collaborating agents — while reusing prior reasoning so no plan starts from scratch.
Optimizable LLM Planning · FutureWeaver, COLM 2026 · Log-Augmented Generation, ICLR 2026
Retrieval that reasons over structured and unstructured data — alignment-oriented retrieval, join-aware multi-table retrieval, generative retrieval, and data-lake search — rooted in a decade of database systems research.
ARM, ACL 2025 · ACL 2024 Outstanding Paper · SIGMOD 2020
AI research platform for SEC disclosures. New York.
Google Research ('21) · Microsoft Research ('20) · Microsoft CISL ('19) · IBM Research Almaden ('18)
Advised by Zachary G. Ives and Dan Roth.
FutureWeaver: Planning Test-Time Compute for Multi-Agent Systems with Modularized Collaboration
Log-Augmented Generation: Scaling Test-Time Reasoning with Reusable Computation
Can We Retrieve Everything All at Once? ARM: An Alignment-Oriented LLM-based Retrieval Method
On Synthetic Data Strategies for Domain-Specific Generative Retrieval
Is Table Retrieval a Solved Problem? Exploring Join-Aware Multi-Table Retrieval
★ Outstanding Paper — Towards Knowledgeable LMs @ ACL 2024
Learning to Decompose and Organize Complex Tasks
Finding Related Tables in Data Lakes for Interactive Data Science
Juneau: Data Lake Management for Jupyter
★ Best Demonstration Award