Projects & Highlights
Agent-driven workflows and applications
Automated Active Inference Research
An automated iterative research loop using Claude Code. Starting from a paper on deep active inference, the agent implemented a baseline algorithm, then autonomously proposed and tested improvements over multiple iterations — pushing performance from a mean reward of 42.6 to 144.5 (>3x), beating DQN on noisy CartPole. All done in a single day with no additional tools.
Computational Chemist Agent
A multi-agent system built with the Claude Agent SDK for automated compound classification on Therapeutics Data Commons (TDCommons) benchmarks. A lead agent orchestrates a computational chemist that writes and executes Python code (RDKit, DeepChem) and dynamically spawns a researcher for web lookups when needed. Improvements are driven by an automated iteration loop: run the benchmark, generate a report, auto-create GitHub issues for improvements, implement fixes in a PR, re-run the benchmark, and merge if results improve. Through this process, outsourcing domain knowledge into specialized classification skills (hERG, Absorption, Metabolism) and adding ML-based predictions, cost dropped from >$1 to ~$0.10 per question and full benchmark accuracy improved from 46% to 80%.
Happy to discuss any of these — feel free to reach out.