Table encoder
[ TODO: describe per-column encoding — numerical scaling, categorical embedding, text embedding, time-series encoder. ]
ARCHITECTURE
Relational Deep Learning, without the PhD
Langsat combines a table encoder, a heterogeneous GNN, and task heads — the same recipe that wins on RelBench benchmarks, exposed as a no-code product.
Three layers
Each layer is automatic but inspectable. Export a model card at any time.
Heterogeneous GNN
[ TODO: describe message passing over the graph defined by your foreign keys — SAGEConv, GATv2, Transformer, GIN. ]
Task heads
[ TODO: describe regression, classification, link prediction, clustering, anomaly detection heads. ]
SCALE
Built for real datasets
- Streaming mode for 20M+ row tables with memory-mapped feature arrays
- Neighbor-sampling loaders for graphs with billions of edges
- Multi-AZ EC2 Spot training with On-Demand fallback — 90%+ reliability
- Per-user training limits, per-organization credit pools, audit-survives-delete
[ TODO: link to blog post on scaling patterns. ]
NEXT
Related
Ready to train your first model?
Sign up free. Train a real model in minutes.