PhD pharmaceutical scientist who builds data and AI systems. Doctoral research taught me to dissect problems nobody had solved before. Five years of customer-facing drug development showed me what regulated environments demand. From wet lab to data warehouse to LLM pipelines, I pick up new domains fast and build solutions that fit the problem and hold up to scrutiny.
Three systems built around one idea: data you can trust. Each vault takes a different kind of difficult data and makes it reproducible, validated, and explorable. All three run live.
Stability data usually live one spreadsheet per project, invisible across studies. AssayVault breaks those silos: 93,000 results across 642 formulations and 24 molecules in one warehouse, every record comparable - with ICH Q1E implemented as executable, versioned regulatory logic.
An exploratory cheminformatics project: ChEMBL, UniChem, and PDBe integrated into a reproducible warehouse - dbt medallion architecture, Dagster orchestration, RDKit fingerprints, activity cliffs, scaffold series, and a UMAP map of chemical space.
A self-hosted intake pipeline for everything you read. Sandboxed Claude Agent SDK sessions turn PDFs, articles, videos and notes into linked, cited wiki pages in an Obsidian vault, with a graph of the whole vault, a browsable library and research that cites its sources.
All demos run on synthetic or public data · the LibrisVault demo serves a read-only vault
Designed, built, and shipped three systems end-to-end, from data model to frontend: AssayVault, SARVault, and LibrisVault. Full lifecycle ownership - requirements, data integration, CI, releases, operation.
End-to-end ML pipelines in Python and SQL, API-based model deployment with CI/CD, and transformer-based generative AI with measurable output evaluation.
Primary scientific contact for 10+ pharma and biotech accounts at a biopharmaceutical CDMO. Developed, implemented, and validated analytical methods, led international cross-functional teams, translated ambiguous client requirements into structured programs, and worked daily in GLP-regulated systems where traceability is mandatory.
Dr. rer. nat., magna cum laude. Metal-organic nanopharmaceuticals for drug delivery; peer-reviewed first-author publications.
Compact, practical articles on data engineering, machine learning, AI systems, and biopharma.
Why LibrisVault trusts the OS sandbox instead of the SDK permission layer. The measurement that settled it.
Choosing a hyperparameter on the same folds that report the score inflates it. Nested cross-validation is what restores the separation.
Textbook nanoparticle behaviour does not transfer to lipid nanoparticles. A cryo-TEM study and one elegant control experiment explain why.
Parts of this site run on a Hetzner box I administer myself. Self-hosting is half hobby, half quality bar: if a system is not worth operating, it is not finished.
LibrisVault exists because I am genuinely obsessed with personal knowledge management - reading widely and keeping what I learn linked, cited, and findable.
Which means the vault doubles as an honest interests section. A few of its current domains:
If your problem involves scientific data or demands deep analysis, I'd love to hear about it.