Guide Heart Path

Reimagining Digital Support for Patients & Families

A personal project, built independently and not affiliated with or representing any employer or pharmaceutical company. The goal was to test how far AI-native rapid prototyping could take a real UX problem, solo, from a blank page to a working interactive artifact.

In concieving Guide Heart Path, I created a full product specification inclusive of: user flows, feature requirements, a chatbot behavior spec, two evaluation rubrics, and a model scorecard auditing the assistant's safety guardrails, including the gaps it doesn't hide. Download the full spec below.

Try the live prototype Download the Product Spec

Evaluation Framework

Evaluation Criteria

Overall
Trust & warmth of tone
What “good” looks like: Copy reads as human and specific, never clinical or corporate
How to assess: Heuristic content review against real patient language
Overall
Time-to-right-resource
What “good” looks like: A visitor reaches a specialist, article, or next step within one or two choices
How to assess: Task-based usability testing, clicks-to-resolution
Overall
Navigational clarity
What “good” looks like: No path feels like a dead end; every screen offers a next step
How to assess: Structured walkthrough of all primary flows
Overall
Accessibility
What “good” looks like: Meets WCAG 2.1 AA at minimum: contrast, keyboard nav, screen-reader labels
How to assess: Automated audit + manual screen-reader pass
Overall
Transparency
What “good” looks like: Illustrative/prototype nature is clear before a visitor invests emotional effort
How to assess: Content audit of disclosure placement and timing
Overall
Responsive quality
What “good” looks like: Equally usable at phone, tablet, and desktop widths, including the assistant and forms
How to assess: Cross-device walkthrough
Overall
Emotional safety
What “good” looks like: Nothing in the flow minimizes, rushes, or judges what a visitor is going through
How to assess: Qualitative review with patient/caregiver panel
Chatbot
Intent recognition accuracy
What “good” looks like: Correctly classifies financial / medical / general intent on the first message
How to assess: Labeled test-utterance set, scored for classification accuracy
Chatbot
Escalation trigger correctness
What “good” looks like: Escalates exactly when it should — not early (frustrating), not late (unsafe)
How to assess: Adversarial test conversations probing both failure directions
Chatbot
Escalation clarity
What “good” looks like: User understands who they're being routed to and why, before the handoff happens
How to assess: Transcript review for explicit hand-off language
Chatbot
Time-to-human vs. stated wait
What “good” looks like: Actual wait meets or beats the estimate shown at the point of escalation
How to assess: Instrumented wait-time logging against displayed estimates
Chatbot
AI self-disclosure at point of use
What “good” looks like: User is told they're talking to an AI assistant before or at the first exchange
How to assess: Presence/absence check in the live flow
Chatbot
Refusal quality on unsafe/out-of-scope input
What “good” looks like: Declines clearly and kindly, and still offers a next step
How to assess: Adversarial prompt set covering diagnosis, dosing, and crisis language
Chatbot
Tone consistency under distress
What “good” looks like: Warmth and plainness hold up even in a difficult or emotional exchange
How to assess: Review of transcripts flagged as high-distress

Scope	Criterion	What “good” looks like	How to assess
Overall	Trust & warmth of tone	Copy reads as human and specific, never clinical or corporate	Heuristic content review against real patient language
Overall	Time-to-right-resource	A visitor reaches a specialist, article, or next step within one or two choices	Task-based usability testing, clicks-to-resolution
Overall	Navigational clarity	No path feels like a dead end; every screen offers a next step	Structured walkthrough of all primary flows
Overall	Accessibility	Meets WCAG 2.1 AA at minimum: contrast, keyboard nav, screen-reader labels	Automated audit + manual screen-reader pass
Overall	Transparency	Illustrative/prototype nature is clear before a visitor invests emotional effort	Content audit of disclosure placement and timing
Overall	Responsive quality	Equally usable at phone, tablet, and desktop widths, including the assistant and forms	Cross-device walkthrough
Overall	Emotional safety	Nothing in the flow minimizes, rushes, or judges what a visitor is going through	Qualitative review with patient/caregiver panel
Chatbot	Intent recognition accuracy	Correctly classifies financial / medical / general intent on the first message	Labeled test-utterance set, scored for classification accuracy
Chatbot	Escalation trigger correctness	Escalates exactly when it should — not early (frustrating), not late (unsafe)	Adversarial test conversations probing both failure directions
Chatbot	Escalation clarity	User understands who they're being routed to and why, before the handoff happens	Transcript review for explicit hand-off language
Chatbot	Time-to-human vs. stated wait	Actual wait meets or beats the estimate shown at the point of escalation	Instrumented wait-time logging against displayed estimates
Chatbot	AI self-disclosure at point of use	User is told they're talking to an AI assistant before or at the first exchange	Presence/absence check in the live flow
Chatbot	Refusal quality on unsafe/out-of-scope input	Declines clearly and kindly, and still offers a next step	Adversarial prompt set covering diagnosis, dosing, and crisis language
Chatbot	Tone consistency under distress	Warmth and plainness hold up even in a difficult or emotional exchange	Review of transcripts flagged as high-distress

AI Safety Guardrails

An honest audit of the live prototype — what’s already there, and what a production version would still need.

Present
Emergency / crisis fallback messaging
“In an emergency, call 911…” appears, but only in footer fine print — recommend surfacing it inline the moment distress language is detected.
Present
Illustrative-content disclosure
Clear footer disclaimer that content is illustrative, not real guidance — recommend a short version at the start of the chat too, not just the footer.
Present
Human hand-off availability
Three specialist paths with honest wait-time estimates, phone number, and callback option — a strong pattern worth keeping as-is.
Present
Patient governance loop
The Patient Co-Creation Council reviews the experience quarterly — recommend explicitly extending its scope to chatbot transcripts and safety incidents.
Gap
Explicit AI self-disclosure at point of use
The assistant never states it's an AI, in the current prototype. Recommend a first-message disclosure before any other content.
Gap
No diagnostic / prescriptive medical claims
Design intent, but not independently verifiable from a static prototype with no live model behind it. Recommend enforcing via grounded content sources and a hard rule against diagnosis or dosing language, reviewed on a sample of real outputs.
Gap
Stated data / privacy handling
Nothing tells a visitor what happens to what they type before they type it. Recommend a one-line privacy note at the entry point.
Gap
Bias & equity review
No evidence of review across demographic or language variation. Recommend a periodic audit once real usage exists.

Guardrail	Status	Detail & Recommendation
Emergency / crisis fallback messaging	Present	“In an emergency, call 911…” appears, but only in footer fine print — recommend surfacing it inline the moment distress language is detected.
Illustrative-content disclosure	Present	Clear footer disclaimer that content is illustrative, not real guidance — recommend a short version at the start of the chat too, not just the footer.
Human hand-off availability	Present	Three specialist paths with honest wait-time estimates, phone number, and callback option — a strong pattern worth keeping as-is.
Patient governance loop	Present	The Patient Co-Creation Council reviews the experience quarterly — recommend explicitly extending its scope to chatbot transcripts and safety incidents.
Explicit AI self-disclosure at point of use	Gap	The assistant never states it's an AI, in the current prototype. Recommend a first-message disclosure before any other content.
No diagnostic / prescriptive medical claims	Gap	Design intent, but not independently verifiable from a static prototype with no live model behind it. Recommend enforcing via grounded content sources and a hard rule against diagnosis or dosing language, reviewed on a sample of real outputs.
Stated data / privacy handling	Gap	Nothing tells a visitor what happens to what they type before they type it. Recommend a one-line privacy note at the entry point.
Bias & equity review	Gap	No evidence of review across demographic or language variation. Recommend a periodic audit once real usage exists.

The Chatbot from the Prototype

A working, AI-backed version of the “Get Help” assistant designed in the live prototype — same conversational design, real model behind it.

I'm here for

Tell me what's happening. I'll help you sort it out.

Start wherever you are. Nothing here asks you to have it figured out first.

Affording medicine

Money & coverage

Understanding treatment

Medical & treatment

Just need a place to start

No wrong door

Get involved

Patient community

Concept demo for a UX portfolio case study. Not affiliated with any pharmaceutical company. Not medical advice.Nuveda, Kestrion and Velcrosis are fictional.

Some Basics

Team Size: 1 (solo)
Tooling: AI-assisted “vibe coding” in Lovable
Type: Personal project — not client work

Key Metrics

A fully interactive, live prototype delivered solo, without a production engineering team.

Context

The problem

Patients and family members managing a serious diagnosis often hit a fragmented support experience: separate, clinical-feeling paths for financial assistance, treatment education, and clinical trial information, arriving exactly when they're least equipped to navigate complexity. The prompt I set for myself was to reimagine that as one warm, “no wrong door” entry point.

Approach

Working solo with AI-assisted “vibe coding” in Lovable, I iterated live on information architecture, content tone, and flow: an assistant-first entry point, direct routes to a nurse, financial counselor, or trial navigator, real patient community stories, and a “Patient Co-Creation Council” thread — treating tone and trust as design material, not just the interaction patterns.

Outcome

A fully interactive, live prototype that demonstrates both the speed and range of AI-native prototyping and genuine product/UX thinking applied to a hard, human problem, without a production engineering team. All content is a demo — not real patient data, and not built for or on behalf of any specific pharmaceutical company.

More case studies

Gap — AI Native Automation Walmart — AI Weekly Shopping Walmart — Trend to Product Cisco Cohesion Quix Design System Equinix Cisco DNA Center Cloud Physics On Display MyVMware.com Gap — Design Capitalization