Projects

Not a scientific assistant but the layer below it — the “when can you trust the result?” spine. A deterministic, zero-dependency pipeline that separates evidence from assumptions, keeps uncertainty explicit, flags missing required data, refuses to accept on gaps (recommends “measure again”), and ships a byte-for-byte reproducible trace an eval harness scores. No LLM in the loop.

The retrieval-grounded assistant living in the corner of every page: it answers questions about my public work from a public corpus, cites sources when it can, and refuses to misattribute others' work. Built on a Node RAG/LLM API.

A realtime voice simulation environment for stress-testing conversational agents — synthetic callers, latency-aware turn-taking, and evaluation hooks.

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.

A deterministic 10-stage remaining-useful-life analysis on public NASA turbofan data, with human decision gates, evidence-backed decision cards, and cited diagnostics.

A controlled QLoRA study comparing Gemma 3 and MedGemma on the same preventive-health task, synthetic data, and evaluation harness.

Enter a SMILES string and compare fixed Ridge, XGBoost, MLP, and fine-tuned MIST-28M predictors on QM9 DFT property labels.

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.

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.

Ask a knee-osteoarthritis question; the answer reports what NICE NG226 and ACR/AF say, re-reading each source so the recommendation strength stays faithful instead of collapsing into one confident answer. A research demo of guideline faithfulness — not medical advice.

An offline, deterministic benchmark that compares random, fixed-grid, and expected-information-gain policies under the same synthetic cyclic-voltammetry world, budget, constraints, noise, and injected instrument failures. It records proposals, validation decisions, observations, QC flags, costs, belief updates, and stopping decisions in auditable ledgers. The evidence is simulated; this is not a laboratory or validated digital-twin claim.

A hands-on lab for training and understanding scientific foundation models — representation learning, property prediction, and model-literacy tooling.