🤝 Top Collaborators
🤓 Latest Submissions
AGENTGUARD AI
AI agents increasingly perform actions through tools, such as booking appointments or updating records. A convincing response does not guarantee that the agent selected the correct action. A request to cancel an appointment could trigger a patient-record update instead. Developers need a clear way to identify and investigate this mismatch before deploying an agent. AgentGuard AI explores this problem through a focused reliability-testing prototype. Its central idea is to compare the action expected from a user's request with the action taken by an agent, then present the difference in a dashboard that supports investigation and correction. The demonstration focuses on a healthcare appointment assistant using synthetic scenarios. In the main example, the user asks, “Cancel my appointment tomorrow.” The expected tool is cancel_appointment, while the demonstrated failure is an update_patient call. This makes an otherwise technical routing error understandable at a glance. The target users are developers, QA teams, and AI integrators building agents for business workflows. In the prototype dashboard, users can review test scenarios, inspect expected and actual tool selections, and examine failure details. The interface also demonstrates the intended progression from identifying a failure to reviewing a fix and rerunning the affected scenario. The project's creative focus is the relationship between user intent and tool action. It brings the request, expected behavior, observed behavior, and failure context into one investigation flow. This is designed to reduce the effort needed to understand what went wrong and what should be checked after a change. The current submission demonstrates this approach through a dashboard and example test data. Full end-to-end execution, automated evidence collection, and verified fix-and-retest automation remain development goals. IBM Bob supports the development of AgentGuard AI as a coding assistant.
27 Sep 2026
Anomaly detection in metadata of documents
Using AI to find mistakes in the metadata of PDF documents. The user can upload a document without metadata, the upload is ingested using unstructured.io. Then AI(Mistral) adds metadata, which is displayed in the UI. The user could now copy that out and use to assign metadata to his documents. The user can then upload another document with the human given metadata, or load a lot of documents with correct metadata from a vectorstore (vectara). The user can also provide his own metadata for the document as a textstring. Mistral then compares the two versions of the metadata and tells the user the differences and similarites between both sets of metadata.
19 Apr 2024
