Custom models & architectures
Designing and adapting compact models around particular tasks, with work on encoders, internal representations and model structure.
Model architecture · Task adaptation · Runtime constraintsAI & models
Small models.
Specific purpose.
Custom model architectures and the tools to configure, train and evaluate them. An engineering approach to AI, with a focus on what a model can actually do.
Current work includes the Aster text encoders and Constellation model tooling. These are development projects; public downloads are not offered on this site yet.
Discuss a modelThe work
Architecture is only
part of the model.
Data, training, runtime configuration and evaluation all shape the result. The aim is to make those choices explicit and useful.
Configuration, training & evaluation
Selecting configurations, preparing training runs and comparing checkpoints on defined tasks. Understanding the limits matters as much as the headline result.
Training recipes · Checkpoint comparison · Task evaluationSoftware for model work
Developing interfaces and utilities for model creation, configuration, training and inspection, so experiments remain understandable and repeatable.
Model libraries · Configuration tools · Experiment workflowsASTER / RESEARCH & DEVELOPMENTText encoders.
Text encoders.
Built and tested.
Aster is Albion’s custom bidirectional text-encoder work. It explores masked-language pretraining and downstream tasks including text classification and named-entity recognition.
The focus is on representations and task performance. Aster is an encoder project, not a chat assistant.
Ask about AsterThe surrounding toolkit
Constellation brings model configuration, model libraries, training and evaluation into a development workspace.
Together, the model and tooling work support a practical loop: configure, train, evaluate, inspect and refine.
text → tokens → encoder → task head
configuration → training → evaluation
configuration → training → evaluation
Development work · Enquiries welcome