Modelle, Dateninfrastruktur und Werkzeuge, die ich gebaut habe, die meisten davon Open Source.
2026
Embedding model for scientific documents, trained with the sentences in which papers cite each other as contrastive signal. ModernBERT-base with up to 8K context; six checkpoints released under MIT.
Up to ~30M training pairs on 4×H200, SciRepEval 66.85 (EMNLP 2026)
PyTorch, ModernBERT, Contrastive learning, DuckDB
2026
Locally deployable infrastructure that joins eight scholarly databases (Semantic Scholar, OpenAlex, Crossref and others) into one SQL interface, with embedding-based alignment of 4,516 topics to 13 ontologies.
293M papers, ~960 GB of Parquet behind 153 DuckDB views
Python, DuckDB, Parquet, Embeddings
2026
Native macOS app that converts PDFs to Markdown and JSON on the device: reading order across columns, tables as tables, maths as LaTeX. Decides per page whether to trust the text layer or run OCR.
Runs a 0.9B OCR model locally in 1.25 GB at 4-bit
Swift, SwiftUI, MLX, Apple Vision
2024
Curated, weighted list of AI concepts, seeded from textbook indices and the Computer Science Ontology, extended with keyphrases from PapersWithCode and filtered with a classifier against OpenAlex abstracts.
10,797 concepts (DRUID 2024)
Python, YAKE, Neo4j, PostgreSQL, Logistic regression
2023
Tool for downloading scientific papers at scale and converting the PDFs to text in parallel, as the data layer for the first PhD experiments.
Python, multiprocessing