SARIEL: a private AI workbench

Software built around real tasks

SARIEL is an internal AI workbench developed by UDS. The implementation includes an operator dashboard, local model selection, project-file retrieval, research tools and utilities for document and spreadsheet work. Research continues alongside that working code.

The useful question is how those capabilities work together: which tool fits the task, what information supports the answer, and which actions need a person to review them.

Find information and keep its source

The project-retrieval implementation reads approved folders, filters out disallowed files and returns relevant text with file and line references. Its results provide evidence for a task while preserving the distinction between retrieved information and a verified conclusion.

The public-web research tools assemble source-labelled material for synthesis. The code includes checks for source references, unavailable providers and unsupported citations. These checks support review; they do not make generated answers infallible.

Practical tools inside the workbench

Spreadsheet tools inspect CSV and Excel structure, identify formulas and prepare draft audit reports. The document utility prepares reviewable text drafts. Local model-routing code chooses between defined roles and fallback paths, while the operator dashboard makes task and review information visible.

Append-only records capture review intentions and evaluated events. The implementation includes checks for duplicate requests, damaged records and changes to the record chain. It provides a trace of recorded actions without treating every proposed action as approved.

Computing is part of the work

The photograph shows the owner-built UDS development workstation. It illustrates the hands-on computing side of the practice; it is not a screenshot of SARIEL software or evidence of model performance.

Current development status

These capabilities are supported by implementation and test code in the internal project. Integration, hardening and broader evaluation remain ongoing; this reference does not present SARIEL as a finished commercial platform or claim that every module is deployed together.

The SARIEL research page explains the wider architectural direction. Client AI systems are scoped and tested for their own information, tasks and operating boundaries.