Digital Media, Arts, and Technology ("DIGIT") @

Research team: Molly Scott Wright and Dr. Elisa Beshero-Bondar

Agentic AI and MCP with Small Language Models

for Digital Humanities Projects

Chat Model

Code Model

MCP Chat

CLI

Sends Change

User interacts with Chat model through terminal

Model proposes change

User approves change

MCP Server

Directories on PC

Mount

Corpus

Backups

Schemas

Logs

Data

No matter how a directory is formatted outside of the container, the files are "translated" into a predesignated structue!

Two Models are Better Than One! The model with file editing permissions is not the same one that the user speaks to - each are trained to be good at what they are focused on.

• The documents we work are often historical and deal with cultural heritage
• Because of this, we need to avoid AI hallucinations.
• A focused, fine-tuned model can work as an effective assistant with less required compute.

Why Small Language Models?

• We prioritize alternatives to Large Language Models.
• We work with SLMs with open weights and transparent training data.
• Should run. completely offline and client-side.
• Secured in a Docker container.
• Minimizes the required tech stack!

The Goal:

To create a secure, local AI Agent application, with Small Language Models (SLMs), to streamline the process of TEI-XML markup encoding and schema validation for scholars.

Check us out on

Codeberg!