Waythread

Inspiration

Seeing my Popop move to a more rural area later in life made the problem of care access feel immediate and personal. Places that once felt nearby became longer trips, transportation became harder to coordinate, and finding the right information often meant searching across provider directories, government programs, transportation services, and individual websites.

As he gets older, he may need more help understanding not only where care is located, but how to get there, what must be confirmed, who needs to be called, and what backup plan exists when the first option does not work. Waythread grew from the idea that people should not have to assemble that entire plan alone.

What it does

Waythread is a web application for non-emergency care-access planning. It helps a person—or someone assisting them—turn a known appointment or named care need into an organized, source-backed coordination plan.

Users can:

plan for themselves or help someone else; start with a known appointment, search for a named type of care, or try a fictional example; compare potential provider organizations; select origin and destination jurisdictions; account for Medicaid, non-Medicaid, mobility, and transportation circumstances; choose whether they expect to drive themselves, use a trusted driver, need a transportation service, or are not yet sure; receive ordered calls, questions, tasks, transportation paths, contingencies, evidence, and a visit handoff; record whether they confirmed or ruled out an option; replan when a transportation path fails; print a concise summary for the appointment and trip.

Waythread is currently limited to an initial area of Virginia. Its county-level transportation coverage is intentionally tiered: some areas have operational coordination rules, some provide curated official links, and others remain clearly blocked until better evidence is available.

It does not diagnose, recommend treatment, book appointments, guarantee transportation, or claim that coverage, eligibility, network status, or availability has been verified.## How we built it I leveraged GPT-5.6 and Codex to build this project. ChatGPT web project folders to research, coordinate, plan, engineer prompts, and keep me organized. Codex to build it, making use of multiple 5.6 family models and reasoning levels, along with plenty of plugins, skills, and every cool little feature I could make use of.

How we built it

I built Waythread as a solo project using GPT-5.6, Codex, and ChatGPT as a coordinated development workflow.

ChatGPT Project folders helped me research the problem, organize evidence, maintain product and trust requirements, plan implementation phases, coordinate tools, engineer prompts, and preserve decisions across a very large project.

Codex served as the primary implementation environment. I used multiple GPT-5.6-family models and reasoning levels for repository analysis, architecture, implementation, testing, accessibility review, security hardening, documentation, and release preparation.

The application uses a deterministic compiler as the source of truth. Facts, contacts, evidence, county rules, action ordering, and safety boundaries remain controlled by validated code. GPT-5.6 Sol was evaluated through a strict structured-output layer for bounded narrative improvements, but model output could not introduce new facts, alter evidence, change transportation classifications, or override deterministic safety rules. Invalid or unsupported output falls back completely to the deterministic plan.

The project was built with Next.js, React, TypeScript, Zod, Vitest, Playwright, and accessibility testing. The final release candidate passed 319 unit tests, 45 component tests, 119 browser tests, a production build, strict type checking, linting, and four accessibility test suites.

Challenges we ran into

The largest challenge was that care access is not a single search problem. Provider identity, insurance questions, transportation eligibility, geographic boundaries, appointment logistics, accessibility needs, and backup planning all come from different sources with different levels of reliability.

Transportation information was especially difficult. A program may serve only a specific county, population, trip type, or insurance category. Some sources were current and operational, while others were incomplete, conditional, outdated, or suitable only as official link-outs. Waythread therefore needed to preserve uncertainty instead of presenting every discovered program as available.

Virginia’s independent cities also created important geographic edge cases. Winchester City is separate from Frederick County, and the City of Fairfax is separate from Fairfax County. Treating postal addresses as jurisdictions could have produced incorrect transportation guidance, so the application requires explicit jurisdiction handling.

Another major challenge was balancing AI flexibility with healthcare-adjacent trust and safety. The product needed to remain useful without diagnosing, recommending treatment, exposing personal information, fabricating contacts, or implying that eligibility and availability had been confirmed.

Finally, the project had to remain usable as a judge-facing demonstration while also supporting deterministic fallbacks when live provider discovery or model access was unavailable. That required extensive context invalidation, fallback behavior, source labeling, privacy controls, and automated testing.

Accomplishments that we're proud of

I am proud that Waythread became more than a static mockup. It is a working, tested care-coordination product with multiple connected journeys.

The project includes:

known-appointment planning; named care-category and provider-organization discovery; Medicaid and non-Medicaid coordination examples; explicit county and independent-city handling; four travel approaches; transportation no-go recording and deterministic replanning; ordered calls and tasks; source evidence and limitations; caregiver and self-planning language; a concise visit handoff; print support; keyboard navigation and responsive layouts; session-only privacy with no accounts, database, or browser persistence; strict emergency and sensitive-data boundaries; deterministic behavior when live services are unavailable.

I am also proud of the engineering discipline behind it. The release candidate passed hundreds of automated tests across unit, component, browser, accessibility, privacy, provenance, and fallback behavior. Waythread keeps important unknowns visible rather than disguising them as certainty.

What we learned

I learned that the hardest part of building an AI product is often not generating an answer—it is defining what the system is allowed to know, what it must never infer, and how it should behave when the evidence is incomplete.

I learned that deterministic code and generative AI work best when they have clearly separated responsibilities. GPT-5.6 can improve explanation and usability, but source-backed facts, contacts, transportation classifications, safety rules, and consequential actions should remain constrained and testable.

I also learned how much implementation quality depends on operational discipline: versioned evidence, explicit approval gates, source dates, fallback paths, context invalidation, privacy reviews, accessibility testing, and documentation are all part of the product—not work that happens after the product is finished.

Most importantly, I learned that care access is fundamentally a coordination problem. Finding a provider is only one step. People also need to understand what to ask, who to call, how to travel, what remains uncertain, and what to do when the first plan fails.

What's next for Waythread

Nationwide and county- or state-specific expansion is the long-term goal, not the current reality.

The next stage would begin with deeper Virginia coverage: reviewing additional county transportation programs, health systems, official provider sources, Medicaid transportation rules, accessibility services, and reliable local backups.

Future versions could add:

more counties and states; additional provider and facility data sources; carefully bounded conversational intake; voice and read-aloud support; multilingual guidance; secure saved plans; caregiver collaboration; richer mapping and trip visualization; provider and transportation source monitoring; clearer eligibility pre-screening questions without making eligibility claims; stronger integrations with public agencies, health systems, and community organizations.

The larger vision is for Waythread to become a trustworthy coordination layer between a care need and a realistic plan for reaching care—while keeping evidence, limitations, privacy, and human confirmation visible at every step.

Built With

• api
• codex
• gpt-5.6
• next.js
• openai
• playwright
• react
• typescript
• zod