SLM development
How can smaller language models serve focused tasks within limited compute and memory budgets? We investigate data selection, architecture, and training choices.
Research that informs engineering decisions.
Our research directions focus on smaller models, adaptation, agent design, and reliability. These are areas of investigation, not claims of published results.
Research directions
Fine-tuning methods
When does model adaptation outperform improvements to retrieval or prompting? We examine data quality, parameter-efficient methods, and held-out evaluation.
Agentic system design
How should context, tools, and human approval interact? We explore bounded workflows, harness design, and failure handling.
Evaluation & reliability
How can evaluation reveal failures before deployment? We study task-based testing, grounding, reproducibility, and regression checks.
Demonstrations & technical artifacts
Our first public artifact explains a design approach. No empirical research findings are published here yet.
An agent workflow with a human decision point
A synthetic purchase request illustrates retrieval, limited tool access, approval, and evaluation.
Read the annotated workflowBring us the problem. We’ll work through the possibilities.
A workflow to improve, a model to adapt, or a prototype ready for its next step.
Discuss your project