I build AI systems that help researchers and clinicians get trusted answers to highly specialized questions.

My work focuses on designing safe, scalable, easy-to-use software for science and healthcare. Most of what I build sits at the intersection of agent systems, mechanistic scientific models, and tool ecosystems that allow AI to interact with real research infrastructure.

Ida is an experimental environment for scientific learning and hypothesis refinement inspired by the philosophy of Imre Lakatos.

Rather than treating models as static predictors, Ida treats them as research programs that evolve over time. Practically speaking, this means incremental edits to a nonlinear dynamical system — typically SBML biological simulations.

Ida is built on top of SciGym, which provides ground-truth biological models and experimental interfaces. This allows hypotheses, interventions, and model edits to be evaluated against a hidden “true” system, turning mechanistic science into a measurable learning environment.

The long-term goal is to explore extensions of the AlphaGo-Zero learning paradigm to scientific discovery. AlphaGo-Zero showed how agents can learn through search in a known environment, and Polu & Sutskever later demonstrated a similar strategy for automated theorem proving. A natural next step is adapting this style of search to the noisy, partially observable setting of nonlinear dynamical systems.

A research program contains:

• Hard Core — foundational assumptions that remain fixed
• Protective Belt — mechanisms that can change as the system learns
• Positive Heuristic — strategies that guide promising directions for refinement

Fit describes how well a model explains existing observations.

In Ida this appears as residual structure — systematic differences between predicted and observed behavior across interventions and time. These residual signatures are treated as signals of where the model needs improvement.

A belt edit is theoretically progressive when it produces new testable predictions that were not implied by the previous model.

Empirical progress occurs when those predictions are confirmed under experiment or simulation.

At the population level, Ida compares research programs by looking at:

• prediction novelty
• anomaly resolution
• confirmation rates
• explanatory scope

The goal is to study how scientific programs evolve, compete, and converge over time.

A surprising amount of scientific work is simply getting tools and environments to run correctly. Daedalus exists to automate that layer.

Daedalus extends ToolUniverse’s composable tool ecosystem through a component called ToolMaker, which converts GitHub repositories into MCP-compatible ToolUniverse tools.

Conversion can use either:

• user-supplied YAML tool definitions, or
• YAML generated automatically from example Jupyter notebooks in the repository

For computational biology repositories, Daedalus looks for explanatory notebooks that demonstrate workflows, allowing the system to extract runnable tools directly from research code.

The goal is to build a growing ecosystem of agent-usable scientific tools.

Current work explores methods for distilling large models into smaller specialized reasoning agents that rely on tools and structured workflows rather than monolithic inference.

Agentic workflows for healthcare and science, Artificial Life, Integration and update of older ML patterns (RFs, SVMs, Logic, Genetic Algorithms) into newer possibilities.