What I'm building
The Lab
My work runs along two long-term lines. One builds things; one proves when they can be trusted — and the second is the method that makes the first credible. Projects come and go; these two rivers don't. Public projects below link to live demos or repositories; early prototypes remain high-level summaries.
Want the full grouped portfolio? All projects →
01 — River
6 AI-assisted molecular discovery
Toward a trustworthy, reproducible system that reasons about molecular properties — from agentic chemistry workflows to foundation-model benchmarks to redox / battery-relevant candidates.
● Live demo
ChemGraph Loop
A real agentic chemistry workflow you can run live
Ask a plain-language question about a small molecule and watch a real ChemGraph agent resolve a molecule, run a simulation, draw a 3D structure, and audit its result against a workflow guard.
Try the live demo ↗
● Live demo MIST Transfer Benchmark
Compare structure-based ML with a fine-tuned molecular foundation model
Enter a SMILES string and compare fixed Ridge, XGBoost, MLP, and fine-tuned MIST-28M predictors on QM9 DFT property labels.
Try the live demo ↗
● Live demo Redox RFB Predictor
A public redox-potential demo that checks molecular identity before prediction
Enter a SMILES or chemical name, inspect PubChem-backed name options and an editable RDKit structure, then confirm it before the fixed Random Forest predicts on RedDB's DFT scale.
Try the live demo ↗
● Live demo Molecular Discovery Workflow
Resolve identity, inspect the structure, then show what evidence is missing
A public, bounded version of a private molecular-discovery prototype: PubChem-backed identity resolution, human structure confirmation, RDKit descriptors, and explicit evidence gaps instead of invented property values.
Try the live demo ↗
● Early independent prototype
● Public repository Property-Aware Molecular Discovery Agent
A forward/reverse molecular-discovery core
An early Python-first core organized around forward and reverse workflows — molecule→property dossier and target-property→candidate discovery — architected around a Query Planner, Skill Registry, and Evidence Objects rather than treating LLM output as scientific truth.
Scientific Foundation Model Lab
Training and understanding scientific foundation models
A hands-on lab for training and understanding scientific foundation models — representation learning, property prediction, and model-literacy tooling.
See on GitHub ↗
02 — River
4 Trustworthy & evaluable scientific AI
The layer below the science: when can you trust a result? Evaluation harnesses, honest limitation boundaries, deterministic reproducibility, and human-in-the-loop gates — the method that makes the discovery work above credible.
● Public repository
Scientific Agent Lab
An evaluation & reproducibility layer for scientific AI
Not a science assistant but the layer below it — the “when can you trust the result?” spine. A deterministic, zero-dependency pipeline that separates evidence from assumptions, flags missing required data, refuses to conclude on gaps, and ships a reproducible, harness-scored trace. No LLM in the loop.
See on GitHub ↗
● Live demo Condition-Monitoring Agent
A deterministic predictive-maintenance pipeline you can run live
A deterministic pipeline agent runs a 10-stage remaining-useful-life analysis on public NASA turbofan data — with human-in-the-loop decision gates, evidence-backed decision cards, and cited diagnostics. No LLM in the default path; every result re-derives from a named artifact.
Try the live demo ↗
● Live demo Preventive Health Model Lab
A QLoRA fine-tuning study you can inspect end-to-end
The same synthetic preventive-health task and QLoRA recipe run on Gemma 3 and its medical sibling MedGemma on one 8GB GPU, scored before and after fine-tuning. The held-out evaluation measures format, safety wording, and numeric grounding. Nonclinical; it does not establish diagnostic performance.
Inspect results ↗
● Live demo Guideline-Faithfulness Chat
Reports clinical-guideline positions while preserving each one's official strength
Ask a knee-osteoarthritis question; it reports what NICE and ACR/AF say, re-reading each source so the recommendation strength stays faithful instead of collapsing two guidelines into one confident answer. A research demo of guideline faithfulness — not medical advice.
Try the live demo ↗