Teaching models to read relational data.
We work on relational foundation models: schema-agnostic pre-training, temporal graph reasoning, and in-context learning over connected tables. Everything that holds ends up in the platform, and much of it ends up here first, in writing.
- Focus
- Relational foundation models
- Methods
- Schema-agnostic pre-training, temporal graph reasoning, and in-context learning over connected tables
- Output
- Peer-reviewed papers, research notes, and model versions that reach production
- With
- University collaborators and the platform team, as one group
Research agenda
What we're working on.
Four questions drive the lab. Where we've argued one properly in writing, the piece is linked.
01Relational foundation models
One model for any schema.
A foundation model for relational data that reads a company's database as it is built (tables, keys, time) and predicts over it, with no model to build for each question.
LLMs vs. RFM →02In-context prediction
Learn from examples, not from training.
Labelled cases from a customer's own history go in as context when a question is asked. Nothing is trained and nothing is stored, and the model still reasons to customers, accounts, or companies it has never seen.
From enterprise data to graph embeddings →03Adaptation
Specialise without merging back.
Self-supervised adaptation on a customer's database, or fine-tuning on one task, inside an isolated workspace, as a one-way fork that never flows back into the base model.
04Temporal reasoning
Read history as it unfolded.
Edges are versioned in time, so the model can ask what a relationship looked like before the default, before the ring matured, before the customer expanded.
Graph foundation models in production →
Series
Graph foundation models, in three parts.
How a company's database becomes a temporal graph, what the model inside does with it, and what changed when it went live.
- Part 1From enterprise data to graph embeddingsAug 3, 2026 · 8 min read
- Part 2Inside the Graph Foundation ModelAug 11, 2026 · 16 min read
- Part 3Graph foundation models in productionAug 21, 2026 · 6 min read
By Jonas Rodrigues
Publications
Everything we've published.
- Learnable matched filter for methane plume segmentation in hyperspectral imageryIEEE Geoscience and Remote Sensing Letters2025
- Topological Neural Networks go Persistent, Equivariant, and ContinuousICML2024
- On the Generalization of Equivariant Graph Neural NetworksICML2024
- dPASP: A probabilistic logic programming environment for neurosymbolic learning and reasoningKR2024
- Leveraging hyperspectral remote sensing and foundation models for greenhouse monitoringKDD · Fragile Earth Workshop2024
- Provably expressive temporal graph networksNeurIPS · arXiv:2209.150592022
- Rethinking pooling in graph neural networksNeurIPS · arXiv:2010.114182020
- Simplifying graph convolutional networksICML · arXiv:1902.071532019