Research, collaborations, product experiments, and technical explorations.
An opinion piece on the chasm between research performance and clinical deployment.
Clinical AI has a deployment problem, and the way we evaluate models today is a large part of why. Three holes in clinical AI evaluation, why AGI won't solve this, the case for specialized agents, and what a new procurement problem demands.
Nature Medicine calls for better evidence. Here is what that looks like.
We trained two sepsis architectures on PhysioNet 2019, then stress-tested them with Pasteur — jitter, blackout, and cohort simulations across two hospitals. The model that won on every standard metric turned out to be the wrong deployment choice.
MMLU multiverse analysis reveals hidden structure
100 representations fused into a cosmic graph reveal 15 topological regions, near-pure semantic islands, and model-specific failure zones that single-number leaderboards miss.
From peer-reviewed research to open-source tool
Pulsar brings the Thema algorithm from our Nature Energy paper to everyone. High-performance Rust implementation with MCP integration for AI-assisted topological data analysis.
In collaboration with UCSB Environmental Studies & Bren School
Coal power is now economically unviable, outpriced by renewables and burdened by rising operational costs. Our Nature Energy paper reveals how strategic, context-specific phaseouts can accelerate this transition equitably.
Jeremy builds a DTI pipeline and wins the Novartis Challenge.
We implemented a GNN over molecular graphs, fused target encodings, and incorporated geometric features to improve binding prediction - with an interactive compound viewer and resources to reproduce.
