Christian Weilbach

Researcher & Systems Builder

I work on structured learning, persistent memory, and runtimes for intelligent systems. My PhD at the University of British Columbia was in Structured Amortized Variational Inference. Before that, a Master's at Heidelberg University in Dictionary Learning with Bayesian GANs, and a Bachelor's implementing online learning for Boltzmann machines on spiking neuron models. I also spent several years studying Philosophy and Cultural Anthropology.

Current work

I am building Datahike — an immutable, versioned data substrate for intelligent systems. Datahike explores how learning, search, and coordination can be treated as long-lived, inspectable processes rather than disposable computations. The ecosystem includes Proximum (versioned vector search), Stratum (columnar SQL with branching), and Scriptum (git-like Apache Lucene).

Among many projects I consulted for Roam Research on their Datalog query backend, and for pol.is on PCA modernization.

Selected research

My doctoral work under Frank Wood (PLAI group) focused on structured amortized variational inference and the role of explicit structure in learning systems. The following papers are representative outcomes of that line of work:

• Graphically Structured Diffusion Models — generative models with explicit structural priors
• Flexible Diffusion Modeling of Long Videos — large-scale generative modeling
• All-in-One Simulation-Based Inference — unifying amortized and online inference
• Daphne: Multi-Pass Compilation of Probabilistic Programs — compiling probabilistic programs into inference graphs
• Decoupling Conflict Resolution with CDVCS — versioned, conflict-aware replication

Full publication list: Google Scholar →

Talks & community

Talks at Google DeepMind London, OpenAI San Francisco, and MILA Montreal. Co-organized the Clojure meetup in Mannheim-Heidelberg and the Machine Learning Meetup Rhein-Neckar.

Perspective

I am interested in systems that support deliberate, programmable organization of work, knowledge, and learning, rather than relying solely on emergent coordination.

Contact

Email: christian@weilbach.name
GitHub: github.com/whilo
LinkedIn