Some recent (both on-going and no longer active) projects in no particular order.

atmos [2025]

Building weather data infrastructure on on-premise servers: multi-TB dataset pipelines, distributed training setup, and a declarative protocol that lets downstream deep learning models specify exactly what geospatiotemporal data they need while all complexity is taken care of.

partijplein.nl [2023]

Human-in-the-loop search engine for the 2023 Dutch election programs that shows exact PDF sources for every claim. Users can ask questions in natural language and get answers with highlighted passages from the original documents, enabling verification and deeper exploration. covered by NOS and AT5. (with Khaled Tamimy and July Jagt).

the scraper factory [late 2023 - early 2024]

Automated API generation for listing websites (rentals, jobs, press releases) using LLM-driven browser automation and code generation. Input a website url, output a fully functioning structured data api. Required solving browser automation, autonomous code generation, and custom parsing libraries end-to-end. Project exposed the critical gap between LLM demo capabilities and production reliability, providing deep insights into failure modes of LLMs in production. (with July Jagt)

debot [late 2023 - early 2024]

Subgrant from Open State Foundation. Finetuned open-source LLMs (Bloom and Alpaca) on Dutch parliamentary debate by processing tens of thousands of parliamentary debate documents. Core engineering challenge was converting pdf documents into high-quality LLM training data while preserving semantic structure, speaker attribution, and contextual relationships. (with Pim Meerdink)

zonopjebakkes.nl [2022 - present]

Passion project that combined open datasets (building shapes, terrain, sun positioning) to compute sunlight exposure for terrace seating across the Netherlands. Started as a personal tool for Amsterdam, went viral nationally with 3500+ user-contributed terraces. Gained national media coverage across television, radio, and print. (with July Jagt and Rob Schilder)

FXR [2020 - early 2024]

Built a data unification framework for music royalty reporting using a custom DSL. Different platforms report identical usage events in incompatible formats, preventing creators from understanding their actual income. Developed abstraction layer and domain-specific language that compressed common transformation operations, enabling rapid pipeline creation to convert arbitrary royalty formats into a unified schema for comparison and analysis.

• Herzog, R., Mediano, P. A. M., Rosas, F. E., Lodder, P., Carhart-Harris, R., Sanz-Perl, Y., & Tagliazucchi, E. (2023). A whole-brain model of the neural entropy increase elicited by psychedelic drugs. Nature Scientific Reports. https://pubmed.ncbi.nlm.nih.gov/37069186/
• van de Pol, I., Lodder, P., van Maanen, L., Steinert-Threlkeld, S., & Szymanik, J. (2021). Quantifiers satisfying semantic universals are simpler. https://escholarship.org/uc/item/1vm445rp
• Bruggeman, J. W., Irie, N., Lodder, P., van Pelt, A. M. M., Koster, J., & Hamer, G. (2020). Tumors Widely Express Hundreds of Embryonic Germline Genes. Cancers, 12(12), 3812. https://doi.org/10.3390/cancers12123812
• Bruggeman, J. W., Koster, J., Lodder, P., Repping, S., & Hamer, G. (2018). Massive expression of germ cell-specific genes is a hallmark of cancer and a potential target for novel treatment development. Oncogene, 37(42), 5694–5700. https://doi.org/10.1038/s41388-018-0357-2