This is a submission for Weekend Challenge: Generosity Edition
What I Built
KindPath is a web app that turns "I don't know where to start" into a concrete next step for someone in an acute crisis: an eviction notice, an empty fridge, a utility shut-off due tomorrow.
You describe your situation in one plain-language sentence (type or speak it, in any language). KindPath returns:
A phased action plan -> "Do right now", "Next 24 hours", "Stabilization & follow-up" with checkboxes, time estimates, and one-tap phone numbers, so there's always exactly one obvious thing to do next.
A know-your-rights notice tailored to the situation (illegal lockouts, utility shut-off moratoriums, expedited SNAP).
Local resources directory links filtered to the user's ZIP code (findhelp.org, 211, HUD Find Shelter, LIHEAP, legal aid) plus a curated national safety net.
A hands-free voice walkthrough that reads the whole plan aloud in a calm, paced voice, highlighting each step as it goes.
Everything is anonymous, there is no account, no sign-in. Plans and history live only in the browser's localStorage; the single situation sentence is the only thing that ever leaves the device. It's offline-first: hotlines, saved plans, and a location-aware fallback plan all keep working with no connection.
I have a background in community support, so when I came across this challenge it intrigued me right from the start. My goal was something an individual could access anonymously, as so many people today need resources but may be in fear of seeking out the services that will help them. I wanted it to be something simple and calming for people.
Demo
Deployed Link (Render): https://kindpath-xz00.onrender.com/
(Note: Hosted on Render's free tier, the first request after idle takes ~30–60s to wake the server, then it's instant)
Code
GitHub Repo:[https://github.com/Flash148/kindpath]
How I Built It
Stack: React 19 + Vite 6 + Tailwind 4 on the front end, a small Express server (server.ts) on the back. The Express process serves the built SPA and proxies every AI call, so API keys never reach the browser.
Planning: Google Gemini. /api/analyze sends the situation plus a normalized location to gemini-3.6-flash and gets back a strict-schema JSON plan (responseSchema constrains urgency, primary need, phased steps, rights notice, and the voice script). Free text like 78701 or Austin, TX is normalized first, bare ZIPs are resolved to city/state so the model works from clean geography.
Designed for grounding, degrades to $0. The key architectural decision: local resources should be real, not hallucinated. So the flow is two calls.(1) a Google Search–grounded research pass that finds actual local organizations with source URLs, (2) a second call that folds that research into the JSON schema, marking anything it can't tie to a source as unverified. Search grounding needs a billed Google Cloud project, so it sits behind an env flag and ships off. On the free tier the app runs on a curated, verified national safety net plus ZIP-filtered directory deep links. Every resource carries a scope (local/state/national) and confidence (verified / AI-suggested) flag, and the UI shows an honest "verify before you rely on this" badge on anything unconfirmed.
Voice: ElevenLabs. The walkthrough streams from ElevenLabs' eleven_multilingual_v2 for a warm, calming read, with a graceful fallback chain: ElevenLabs -> Gemini TTS -> the browser's built-in speechSynthesis. It works with no key at all, just with a more robotic voice.
Resilience over cleverness. The free Gemini tier throws frequent 503 high demand spikes, so every model call goes through retry-with-backoff and a fallback model. If the AI path fails entirely, the server returns a location-aware offline plan built from the same safety-net data, the app never dead-ends on a blank screen, which matters for this use case.
Trauma-informed UX. Calming note before the checklist. One obvious next action. High-contrast mode, three font sizes, per-step "read aloud", large tap targets, and a one-tap "quick exit" that redirects away from the page.
Prize Categories
- Best Use of Google AI: Gemini generates the entire structured action plan and rights guidance; the two-call architecture is built around Search grounding for real local resources.
- Best Use of ElevenLabs: streaming eleven_multilingual_v2 voice walkthrough for hands-free, calming guidance, with a full fallback chain.
Screenshot:
Top comments (0)