Evergreen

Inspiration

Caregivers of people living with dementia are often told that familiar music, old films and favourite foods can help a good moment happen. What nobody tells them is what to try next for this particular person. Evergreen starts from what someone has always loved and turns it into a short, gentle list of ideas for the week, so a caregiver is not starting from a blank page.

What it does

A caregiver enters three to six loved things (a singer, a film, a show), the decade the person was young, a town or city, anything to avoid, and any comfort needs such as hearing or mobility. Evergreen asks Qloo to match each one, and the caregiver confirms the right match. Nothing is picked for them.

Evergreen then builds a weekly Memory Kit: Listen, Watch, Go and Talk, plus Taste when Qloo returns food places. Every idea has a 10 minute and a 30 minute version.

Every idea is labelled by where it came from:

• From Qloo (solid green edge): it came from Qloo's results. "Why this?" shows Qloo's ID, when it was fetched, and which loved things it was linked to.
• AI suggestion (dashed purple edge): the language model proposed it.

The label uses words and shape, not only colour. The caregiver can react Lit up, Neutral or Upset on each idea. Upset ideas never come back, and Lit up ideas guide the next round (up to three refreshes). The kit can be compared side by side with a plain AI kit written without Qloo, with no verdict, and printed as one page.

How we built it

• Qloo: /search to resolve each loved thing to a real entity, then /v2/insights for artists, films, TV and places, using the loved things as signals, with explainability turned on.
• Language model: an open model (gpt-oss-120b) on Groq's free tier writes the kit from a compact evidence set. It may only cite evidence IDs that exist.
• Guards in plain code: they reject items that cite missing or mismatched evidence, claim to be from Qloo when they are not, hit the avoid list, or contain medical claims. If Qloo cannot fill a part, the kit says so instead of padding it.
• Privacy by design: no database. The kit lives in the browser tab, and API keys stay on the server. The caregiver's nickname never leaves the browser.
• Stack: Next.js 16, React, TypeScript, Tailwind CSS and Zod. About 570 unit tests and 40+ Playwright browser tests (desktop and phone, with axe accessibility checks).

Challenges we ran into

• Fuzzy matching: Qloo's search is fuzzy. In our own test set, "Umm Kulthum" resolved to an unrelated DJ. One wrong match would quietly steer the whole kit, which is why the caregiver confirms every match and our test tooling refuses a non-matching name.
• Qloo's coverage: era filtering works for films and TV only, region steers places only, and songs and recipes are not Qloo entity types, so those are labelled AI suggestions.
• Model reliability: models invent specifics, so we check every item instead of trusting it. The free Groq tier also limits how fast we can generate.
• A sensitive domain: we avoid medical claims, and the product says plainly that it is not medical advice.
• Hardening: independent security reviews found real problems, such as rate-limit bypasses and an AI item that could claim a Qloo source. We fixed them and added regression tests.

Accomplishments that we're proud of

• Every idea can be traced to its source, and the product never presents a model's guess as Qloo's.
• It works end to end on live data in about fifteen seconds.
• It is accessible: keyboard use, visible focus, and automated axe checks pass on every main screen.
• We built a blinded comparison toolkit so the "does Qloo actually help?" question can be answered by human raters, not by us.

What we learned

• Showing the evidence matters more than sounding confident.
• Deterministic checks around a model catch more problems than a better prompt does.
• Matching the right entity is the foundation, because everything downstream depends on it.
• Honest limits are a feature: this is a prototype, it has not been tested with caregivers, and we have not proven that Qloo beats a plain AI model.

What's next for Evergreen

• Run the blinded comparison with real raters, ideally caregivers and activity coordinators, and publish the result whichever way it goes.
• Get feedback from caregivers and care-home activity coordinators before claiming any benefit.
• Add a shared store for rate limits, and deeper era control.
• Shareable kits for care teams.

Built With

• api
• axe-core
• css
• github-action
• gpt-oss
• groq
• next.js
• node.js
• playwright
• qloo
• react
• tailwind
• typescript
• vitest
• zod