Building the structural foundation for Artificial Chemical Intelligence.
I am executing a concrete roadmap to evolve AI from memorizing chemical history to deriving chemical logic—designing generative systems that master the grammar of reactivity, and building the infrastructure required to validate them rigorously, under the isChemist Protocol.
We are attempting to solve chemistry's quantitative problems (toxicity, binding) before we have mastered its grammar. I build models that learn the causal logic of reactivity, not just static property correlations.
Science is infrastructure-constrained. Bad software forces the mind out of flow and into debugging. I treat developer experience (DX) as a scientific prerequisite, building tools that make validation invisible.
Proof of the "Infrastructure" thesis. I build and maintain full-stack platforms that solve the evaluation and deployment bottlenecks in our field.
A public leaderboard and visualization platform for AI retrosynthesis. Compare models side-by-side, inspect predicted routes, and see which ones generate sound chemistry versus chemical fiction.
The engine that powers SynthArena. A python toolkit that solves the babel of formats in AI retrosynthesis, providing a universal translation layer and a reproducible, cryptographically-verifiable pipeline for benchmarking any model.
A full-stack application that hosts computational chemistry models from 5 publications, making them accessible to the public. Has served over 130 users and processed more than 1,000 prediction tasks.
Operationalizing the "Structure" thesis. I build systems that master chemical structure (Retrosynthesis), navigate chemical space (Active Learning), and compute properties efficiently (Quantum).
Moving beyond search. Developing Project Ariadne and DirectMultiStep to generate synthesis routes as coherent linguistic sequences, capturing the global grammar of reactivity.
Navigating the structure of chemical space. Designing property-based acquisition functions (ChemSpaceAL) that find novel drug candidates with 10x greater data efficiency.
Opening the black box. Investigating whether models like Ariadne are learning chemical principles or simply memorizing dataset biases.
Scalable precision. Creating composite methods to predict quantum properties with gold-standard accuracy at a fraction of the computational cost.
A breakdown of Procrustean Bed for AI-Driven Retrosynthesis: A Unified Framework for Reproducible Evaluation
a deep dive into how the metrics for ai chemical synthesis are broken, rewarding models for "solving" routes with impossible, hallucinatory chemistry. we present the data and the open-source tools to fix it.
A breakdown of ChemSpaceAL: An Efficient Active Learning Methodology Applied to Protein-Specific Molecular Generation
we built an ai steering system for molecular design that efficiently guides a generative model to invent potent, protein-specific molecules, even rediscovering existing drugs from scratch
A breakdown of MP2-Based Composite Extrapolation Schemes Can Predict Core-Ionization Energies for First-Row Elements with Coupled-Cluster Level Accuracy
we developed a computational shortcut that predicts quantum properties with gold-standard accuracy at a fraction of the cost, making high-precision modeling accessible for larger molecules.
Cursor is dying because cost-optimization forces models into tunnel vision. RAG agents fail because they only see what they search for. The superior workflow for 2026 is massive context windows (gemini 2.5 pro) and manual control. Stop letting agents hide code from the model.
science is infrastructure-constrained, not ideas-constrained. an argument on why better tooling, not smarter models, is the missing link in the AI for Science revolution.
We will see why chemistry could not exist as a science without Born Oppenheimer approximation and what career paths that approximation opens.