Case Study · Generative AI · Civic Infrastructure
Making healthcare policy legible — so participation is possible.
A generative AI beta exploring how dense healthcare policy documents could become more understandable, more actionable, and more open to public participation.
- Role
- Founder · Product Lead
- Project
- CG Advocacy Platform
- Period
- 2023 — 2024
- Context
- NSF I-Corps customer discovery · Microsoft AI Hackathon · Independent applied AI prototype
- Stack
- GPT-4 · Retrieval-augmented generation · NLP · Policy document parsing · Persona-based explanation flows
- Research Focus
- Could AI help people understand the healthcare policies that shape their care — without replacing judgment, consent, or civic voice?
Thesis
AI is not the product. Participation is.
Healthcare policy affects patients, clinicians, caregivers, hospitals, rural communities, and public agencies. But the documents are often written for legal, regulatory, and institutional audiences.
The people most affected are not always the people most able to respond.
This project explored a civic participation layer for healthcare policy — one that could translate dense regulatory language into clearer explanations, help people understand what a rule might mean for their context, and support more structured public comment.
Not automation for its own sake. A bridge.
The Problem
Healthcare policy is not equally legible. That matters.
Public comment is technically open.
But access is not the same as participation.
Healthcare policy documents from agencies such as CMS, FDA, and ONC are long, dense, and difficult to interpret without regulatory fluency. Comment windows are time-limited. Submission processes can be intimidating. And the people closest to lived impact may have the least support turning experience into formal feedback.
The result is a participation gap. The system hears from institutions. It hears less often from patients, rural communities, caregivers, frontline staff, and smaller organizations carrying the consequences of policy change.
Three Barriers
What keeps participation out of reach.
- 01
Complexity
Regulatory language creates friction before participation begins.
People may have something important to say, but no clear way to understand what part of the policy affects them.
- 02
Asymmetry
Organizations with policy teams have more capacity to respond. Patients, advocates, rural stakeholders, and small clinical operators often do not.
The gap is not interest. It is infrastructure.
- 03
Signal Loss
Even when feedback is gathered, it can be fragmented, emotional, duplicative, or hard to synthesize.
That does not make it less valid. It means the system needs better ways to preserve meaning while improving structure.
The Solution Direction
A translation layer between policy documents and public voice.
The CG Advocacy Platform beta explored three functions:
- Explain
- Translate dense policy language into plain-language, persona-specific guidance.
- Guide
- Help users understand which parts of a policy may matter to their role, community, or care context.
- Structure
- Support users in drafting public comments that preserve their voice while making feedback easier to review, group, and act on.
The goal was not to replace policy expertise. The goal was to make participation less dependent on already having it.
Product Architecture
Three pillars shaped the prototype.
- 01
Document Simplification
Used retrieval-augmented generation to parse long policy documents, retrieve relevant source passages, summarize key sections, and generate plain-language explanations.
Users could ask questions in natural language and see responses grounded in source material.
The design priority was provenance. Not just an answer. A traceable answer.
- 02
Persona-Based Experience
Designed guided flows for different stakeholder types: patients, advocates, hospital administrators, policy researchers, and health economists.
Each persona needed a different explanation of the same policy. Not because the facts changed. Because the implications did.
- 03
Structured Comment Support
Explored how user input could be gathered, clarified, grouped, and formatted into public-comment drafts.
The user stayed in control. The platform helped turn concern into language.
Voice first. Automation second.
Process
How the beta came together.
- 01
Discovery
Conducted 25+ customer discovery interviews through NSF I-Corps and related stakeholder outreach.
Participants included hospital administrators, rural health advocates, policy researchers, economists, and healthcare operators.
The discovery question was simple: where does policy become too dense to act on?
- 02
Journey Mapping
Mapped the path from policy release to interpretation, internal discussion, public comment, and implementation planning.
The friction was not only document length. It was timing, role clarity, confidence, and translation.
- 03
Prototype Design
Built persona-based flows for policy explanation, question asking, comment drafting, and feedback structuring.
The prototype tested whether users could move from confusion to informed participation with less effort.
- 04
AI System Design
Developed a retrieval-based architecture for document parsing, source-grounded Q&A, keyword extraction, and thematic clustering.
The model was treated as one component inside a larger trust system. Source material mattered. User control mattered. Revision mattered.
- 05
Beta Testing
Tested early flows with healthcare stakeholders and policy-adjacent users.
Feedback shaped the interface language, persona logic, transparency requirements, and guardrails around generated comment drafts.
What the Prototype Tested
The beta focused on four questions.
- Could users understand a policy faster?
- Could explanations be tailored without distorting meaning?
- Could user feedback become more structured while preserving voice?
- Could source traceability increase trust?
The early signal was encouraging.
Users valued speed, but they trusted the system more when it showed where an explanation came from and gave them control over what was submitted.
Early Results
Directional findings from beta testing and stakeholder feedback.
- Comprehension
- Users were able to move from dense policy text to clearer role-specific understanding more quickly.
- Access
- Persona-guided flows helped non-expert users identify which sections of a policy were relevant to their context.
- Response Time
- Structured explanation and drafting flows reduced the time required to prepare a comment or internal response.
- Feedback Quality
- Comment support helped turn broad concern into more organized, submission-ready language.
- Trust
- Users responded positively to source traceability, edit control, and transparency around what the system generated.
Stakeholder Views
Different users brought different questions to the same document.
Health Economist
What changes in cost, reimbursement, burden, or regional impact?
The platform helped surface fiscal implications and translate them into clearer policy questions.
Rural Policy Maker
What does this mean for communities with fewer resources?
The platform helped organize rural-specific concerns so they could become clearer policy signal.
Hospital Administrator
What will this require operationally?
The platform helped identify implementation implications, funding gaps, staffing pressure, and compliance questions.
Technical Components
Built for legibility and control.
- 01
Prompt + Retrieval Design
Developed prompt templates and retrieval logic for persona-specific explanation, Q&A, and comment support.
Outputs were designed for review. Not blind acceptance.
- 02
Document Parsing
Structured dense policy text into retrievable sections so generated responses could remain connected to source material.
The system needed to show its work.
- 03
Thematic Clustering
Explored semantic clustering to group public feedback into themes, concerns, and coverage gaps.
The goal was to preserve civic voice while making patterns easier to see.
- 04
Transparency Log
Designed a traceability concept showing what the system generated, what the user edited, and what the user chose to submit.
Consent had to remain visible.
Strategic Insights
What this taught me about applied AI.
- 01
Trust mattered more than fluency.
A polished answer was not enough. Users wanted to know where the answer came from, what had been changed, and whether they could safely disagree with it.
In public infrastructure, provenance is not a feature. It is the condition for use.
- 02
Co-creation shaped the product.
Economists, advocates, hospital administrators, and patients did not ask the same questions. They did not use the same vocabulary.
The product improved when those differences shaped the interface instead of being flattened by the model.
- 03
AI had to earn its place.
The strongest parts of the platform were not the model outputs alone. They were the surrounding design decisions: what source is retrieved, what context is shown, what the user controls, and what happens before anything is submitted.
The intelligence was not only technical. It was civic.
Reference
"Christine's standout quality is her constant curiosity, which she demonstrated in recent projects leveraging generative AI to drive innovation. Her contributions to projects involving CapsicoHealth's AI-driven analytics products helped demonstrate the value of these tools to stakeholders. Her ability to articulate the business and clinical implications of cutting-edge technologies makes her an invaluable collaborator in aligning AI capabilities with healthcare objectives."
What This Changed
This project clarified my stance on AI.
The question is not whether AI can generate. It can.
The question is whether it can support trust, agency, and participation inside systems where people already feel excluded or overpowered.
Healthcare does not need more opaque automation. It needs better civic interfaces.
- —Interfaces that make complexity legible.
- —Interfaces that preserve voice.
- —Interfaces that help people act before decisions are finalized without them.
Collaborate
Applied AI that earns its place inside systems people already have to trust.
Open to senior strategy, applied AI, and public-interest technology roles at the intersection of healthcare, civic infrastructure, policy translation, and human-centered system design.
← Previous case
Novant Health — Rural Behavioral Health Service Design
Next case →
Healthcare Control Center — Decision Environment Concept