Claribel — USDA GenAI Chatbot

USDA Farm Loan Programs staff navigate 2,000–3,000 pages of policy handbooks to answer questions from farmers applying for loans. New customer service representatives relied on Ctrl+F or asking a more experienced colleague. Seasoned loan officials needed deeper cross-referencing across multiple policy areas. Leadership wanted to deploy a GenAI chatbot that could surface policy answers with citations to source material — but the team had no UI design and no UX perspective on how to make AI-generated responses trustworthy and actionable in a government context.

I started by interviewing the product owner to identify the three core user types: customer service reps looking for quick answers while on the phone with a farmer, subject matter expert loan officials doing deep policy research across multiple handbooks, and executives who needed a mobile-optimized view to look up information while in meetings. Each persona had fundamentally different needs — the CSR needs a short, actionable answer fast; the SME needs comprehensive source material and cross-references; the executive needs the same information delivered cleanly on a phone screen.

The stakeholder's three priorities were clear: access to ground truth source documents was non-negotiable, the system needed to help users construct accurate prompts, and there had to be a structured feedback mechanism so the team could improve the AI's responses over time.

I designed a personality system that adapted the chatbot's behavior to the user's context. A "Customer Service Coach" mode delivers short answers with suggestions for phone or email follow-up. A "Technical Research Assistant" mode provides longer, more detailed responses with additional references and diagrams. Users could also filter which policy documents the AI draws from — critical because the AI model was getting confused between loan making and loan servicing questions when the full corpus was included.

For the citation experience, I designed a reference system where numbered citations in the response link to a split-screen document viewer showing the highlighted source passage alongside the AI's answer. This was the product owner's highest priority — staff needed to verify AI responses against the actual handbook language and cite those sources in official correspondence like adverse decision letters.

I also designed the feedback mechanism with structured categories (factually correct/incorrect, contextually relevant/irrelevant, easy/difficult to understand, practical/impractical) plus an optional free-text field, giving the development team actionable data to improve the model.

Since the development team didn't have Figma access, I delivered fully dimensioned specs — redlined screenshots with spacing, font sizes, weights, and corner radii annotated — so engineers could implement without ambiguity.

Designing for AI-generated content requires different trust patterns than traditional interfaces. In a government context, the AI's answer isn't the product — the verifiable source citation is. Everything in the UI had to reinforce that the chatbot is a tool for finding information faster, not a replacement for policy expertise. The personality and filtering systems were my solution to a prompt engineering problem that was showing up as a UX problem — the AI was giving confused answers not because the model was bad, but because users had no way to scope their questions.