Wisconsin DOR

Wisconsin property tax guidance spans statutes, administrative rules, case law, assessment manuals, government publications, and professional standards. The hard part is not generating an answer. It is finding the right authority, preserving legal context, and showing exactly where the answer came from.

The prototype lets users ask natural-language questions and receive streamed answers grounded in public, directly linked sources.

As an engineer on the DxHub team, I worked across the retrieval system, ingestion pipeline, cloud infrastructure, and end-user experience. My contribution centered on making the system’s evidence more accurate, observable, and useful.

• Developed authority-aware GraphRAG retrieval and structural document browsing.
• Built source-linked citation cards and page-level citation navigation.
• Improved statute, administrative rule, and WPAM chunking quality.
• Added retrieval traces and an admin activity surface for evaluation.
• Consolidated the ingestion workflow into managed Fargate phases.

An authenticated question becomes an EventBridge event. A Lambda then runs a Claude tool-use loop across FAQ search and Neptune graph tools. The agent can perform vector search, traverse legal relationships, inspect document structure, and determine when it has enough evidence before streaming an answer and its sources over WebSocket.

The knowledge graph models a nine-level authority hierarchy, from constitutional and statutory sources through case law, administrative rules, manuals, FAQs, and standards. Relationships such as cites, implements, supersedes, and conflicts with help retrieval preserve context that a flat vector index would lose.

The resulting open-source prototype connects a conversational interface to an inspectable evidence trail. Users can follow citations to public sources, while the team can investigate retrieval behavior, feedback, and document quality without treating the model as a black box.