Field Notes
Working notes from building scientific AI systems — agents, RAG, evaluation, and the open problems in between.
A molecular workflow should show what it does not know
A project note on a public molecular-discovery demo: resolve identity, confirm structure, calculate structure descriptors, and make evidence gaps explicit instead of inventing properties.
AI for ScienceA redox predictor needs a structure checkpoint
A project note on the SciencesLoop Redox RFB Predictor: why name resolution, editable structures, and explicit confirmation come before an electrochemical prediction.
AI for ScienceCan a fine-tuned molecular foundation model match engineered traditional models?
An independent QM9 case study comparing a released fine-tuned MIST checkpoint with Ridge, XGBoost, MLP, and two carefully separated ensembles.
AI for SciencePaperQA2 is a useful reminder for scientific RAG: leave an evidence trail
A SciencesLoop technical note on PaperQA2, citation precision, and reviewable scientific RAG artifacts.
AI for ScienceChemGraph is a better first test for scientific agents
A hands-on SciencesLoop note from running ChemGraph locally, finding workflow failures, and checking whether the agent showed enough evidence to trust the result.
AI for ScienceAgentic Scientific Workflows: From Search to Traceable Decisions
A working model for scientific agents that retrieve evidence, plan the next step, preserve uncertainty, and leave a decision trail.
AI agentsA Public-Safe Site Agent as a Portfolio Interface
Why a personal technical site can be more useful when the homepage becomes a small, cited, public-safe agent.
site agentScientific RAG Needs Evaluation Before Confidence
RAG systems for technical work should be judged by retrieval quality, citation faithfulness, and decision usefulness, not only fluent answers.
RAGEvery note starts from public sources and personal testing — evidence first, interpretation second.