Inspiration: Guide dogs help people with disabilities move through a world that wasn't built for them. They lead, warn about obstacles ahead, and let their handler make the decisions. We saw the same problem online, especially on patient portals, where people book appointments, refill medicines, and read test results. These sites are often full of tiny buttons, unlabeled forms, and hidden timers that sign you out and erase your work. So we built a guide dog for medical websites.
What it does: Pawthfinder is a browser extension that sits on top of a patient portal and guides people through it, led by a friendly animated guide dog, Barnaby. Here are some main features that are of most importance to users: Asks, doesn't assume - When you start, it asks what helps you: voice control, bigger text, plain words, keyboard use, and more. The site adapts instantly. Leads the way - Say or type a goal like "Book a follow-up with my cardiologist." Barnaby turns it into a list of steps across several pages and walks you through each one. Translates - It puts jargon into plain words, such as "Encounter Mgmt" becoming "Appointments", at three reading levels, and can read pages aloud. “Sniffs” ahead - It warns you about barriers before you reach them, like mouse-only robot checks, fine-print payment agreements, and phone-only steps. Protects your work - It spots hidden sign-out timers, keeps you signed in, and saves your answers. Keeps you in control - It never fills in passwords and never ticks a consent box for you. Stays private - Everything stays on your device, and it works offline.
Challenges we ran into: We had no real patient portal we could test on, since real ones hold private health data. So we used Claude to build MyHealthLink, a mock portal with the barriers disabled patients really run into, built in on purpose: unlabeled fields, tiny mouse-only buttons, a mouse-only robot check, and a hidden four-minute sign-out. That gave us a realistic "before" to compare against Pawthfinder's "after". We also had to design for many disabilities at once without forcing anyone into one category. That led us to replace fixed presets with a "what helps you?" picker.
Accomplishments we're proud of: We built AI test agents that act as disabled users: The agents found real problems a sighted mouse user would never notice, like keyboard focus getting lost after closing a notice. That gave us a prioritized list of fixes. What we learned: AI works best with focused jobs, not one giant prompt. We split the work across specialist agents: one for performance, one for visual design, one to test as disabled users, and a manager agent that checks everyone's work against our goals and stops agents from overwriting each other. We learned to give each agent clear rules about which files it owns. We also learned that accessibility has to win whenever it conflicts with looks or speed.
What's next for Pawthfinder: More portals - Use an AI model to understand any medical website, not just our demo, while keeping health data on the device. Connect to care - Link Pawthfinder to real portal systems so people can reach their doctors, send accommodation requests, and keep appointments in sync. Test with real people - Run the app with disabled users, screen readers, and voice-control software, not just simulations.
Log in or sign up for Devpost to join the conversation.