Projects
Large language models and agentic assistants for physics research. Fine-tuning, local inference on AIP GPU clusters, and AI agents that execute real scientific workflows end to end — including driving REANA research platforms from natural language. Co-developed with Dr. T. Tong at AIP.
A 400 million-star spectroscopic catalogue of the Milky Way. I own the end-to-end data pipeline: float32 schema design, full-catalogue Parquet joins (210 GB) on HPC, the plotting engine, and the publication figure suite — plus the science work on the bulge, bar and disc populations it enables.
Interactive 3-D particle and cosmological-simulation viewer (OpenGL/GLU/GLEW). B-tree spatial acceleration for tens of millions of particles, HDF5 snapshot reader with on-the-fly disk cache, camera/fly-through navigation, particle-selection tooling, PCA and histogram diagnostics. The tool behind visual exploration of MareNostrum-class simulations; open source since 2005.
A virtual reality tour of the Leibniz Institute for Astrophysics Potsdam — vr.aip.de. 360° photography of the Telegrafenberg campus and observatory halls turned into an interactive browser/VR tour (Panotour pipeline) for public outreach and campus visits.
Research infrastructure for Digital Research Products in particle physics and astrophysics under the PUNCH4NFDI project — reproducible, FAIR (Findable, Accessible, Interoperable, Reusable) analysis workflows as citable products. I contribute to the P4N hub instance (drphub-p4n.aip.de), the REANA-based execution backend, and the compute/storage layer. First author of the WCEDMVO maturity model (with E. Sacchi and H. Enke) that gives DRP-Hub its computed L0–L4 levels.
▸ DRP-Hub deep dive assistant · WCEDMVO 7-component model · L0–L4 maturity · FAIR crosswalk · live registry
The DRP AI Assistant
The hub ships an interactive DRP Assistant — a chat agent over the registry API that helps researchers discover, describe, clone and mature Digital Research Products from natural language. It streams responses (SSE), surfaces the tool calls it makes (each answer can be traced back to the originating API call), keeps per-user chat history, and can be extended with a configurable terminal proxy for repository and execution actions. Admins configure the model endpoint and the assistant's behavior in settings; the p4n instance is pre-production (BETA) and not a citable corpus.
WCEDMVO — a computed maturity model
Registry methodology that separates an evolving product from the evidence backing its current assessment. A product keeps one identity while its component references are attached, replaced or detached; maturity is then computed per record revision from explicit, inspectable gates — a level plus the exact unmet criteria, not a black-box score.
- W · Workflowmanifest, step order, parameters, entry command
- C · Coderepository with release or commit reference
- E · Environmentdependency specification, container image reference
- D · Datainput references, versions, checksums, access conditions
- M · Metadatadescription, creators, citation, license
- V · Validation & provenancescoped checks, outcomes, recorded runs
- O · Outputstables, figures, reference results, tolerances
The L0–L4 maturity ladder
Gates are cumulative: failing L2 (e.g. removing the environment image) drops the product to L1, and all higher gates fall with it. A level describes local policy conformance — it is a navigation aid, not a certificate of scientific validity or full FAIR compliance.
FAIR in DRP-Hub — and where it stops
The ladder and the FAIR principles answer different questions: FAIR concerns properties of data and metadata (persistent identity, semantic interoperability, stewardship); the ladder orders selected local obligations. Each FAIR principle maps to a supporting mechanism — DOI/archive (F1, L4), description + citation + license (F2, L1–L2), HTTPS API and artifact links (A1/A1.1/A1.2), separate registry record for metadata retention (A2), structured metadata with possible RO-Crate export (I1), component links (I3) and license + provenance + workflow fields (R1.1–R1.3) — while the remaining evidence obligations stay explicit gaps. In short: DRP-Hub makes FAIR trackable field by field; it does not by itself certify FAIR compliance.
Live registry (p4n snapshot, 2026-09)
The registry is in pre-production: products are registered and REANA-attached, but almost none have passed the L1 documentation gates yet — the gap the maturity ladder is designed to close, one remediation at a time. The v1 API adds field projection, ETag concurrency, HATEOAS links, SSE events, dry-run on mutations, and idempotency keys; JWT + service-token auth.
A human-centric AI assistance system for astronomy: a curated, human-reviewed skill library that teaches AI agents reproducible astrophysics workflows — catalogues, simulations, instrumentation, data pipelines.
Compute and storage infrastructure for the German National Research Data Infrastructure (NFDI): scalable, reliable HPC services for scientific data — the substrate that DRP-Hub and the AIP data services run on.