That is a lot to carry, especially on your own. Here is a place to start.
Kedara Guide is AI care coordination for families: a caregiving assistant that answers questions, finds local support, and helps plan care. It is live, free, and requires no account. This case follows one question through the product, a daughter asking how to look after her father at home.
My role
I was the AI Product Designer on the team, and the product's UX was mine: the conversation design, the accessibility work, the guardrails, and the voice interface. The persona, the component system, the sourcing pattern, and the crisis boundaries in the sections below are decisions I made and defend here.
◇ marks a waypoint. Click one and the stage beside you changes.
Kedara's users are caring for an ageing parent: 40 to 65, working full time, often in a different city, splitting the work with siblings, with no medical, legal, or eldercare training. Some users are the elders themselves. Some are working in a second language.
Every answer leads to an action in a real house with an elderly person in it. Medication, mobility, money, and in some cases moving someone out of their home. And, a pattern straight from our user interviews: caregivers usually ask at the end of a long day, often right after a difficult phone call.
For a reader in that state, the default shape of an AI answer is a design failure three times over:
The information is fine. The interface is what fails. That made it a design problem before it was ever a content problem.
The direction did not come from a literature review. It came from our own interviews and usability tests with caregivers, run as a team across every version. : the assistant sounded too formal and cold, and the screens held too much at once. Those two notes drove the persona rewrite and the component system that follow.
Only then did we take the notes to the published record, and it said the same things with citations attached. Meeting accessibility standards does not make an answer usable; contrast ratios and font sizes were the entry requirement, and the shape of the answer was the harder problem. in a 2025 study built on 20 contextual interviews, older adults trusted chatbot answers while telling the researchers they did not know where the answers came from (Enam, Murmu & Dixon, Intl. Journal of Human-Computer Interaction).
Only 25% of Americans aged 50 to 64, and 10% of those 65 and older, have ever used ChatGPT (Pew Research Center, 2025). The first exchange carries more weight than it would with a younger audience.
Insight: Clinical walls, hedges that never commit, and confident summaries with no source: each fails a stressed 40+ caregiver in a different way, and published guidance for AI serving older adults asks for clear instruction and empathetic language together.
Kedara began voice-first. A microphone, prompt pills, answers read aloud. Spoken answers vanish: nothing to re-read at midnight, nothing to act on. Text arrived in the next version.
Then formatting depended on the model. Some answers gained bold text and line breaks; structure was whatever the model produced. The same question returned a readable answer one time and a wall of text the next.
Then structure arrived without sourcing. Answers became scannable, but claims still had no origin and the line between information and advice went unstated. Readable, and still unsafe to act on. The feedback that ended each version came from the sessions: too formal, too cold, too much on screen. The screens beside you are the real ones; click any of them to see it full size.
The first question I asked was what kind of helper a caregiver in that state actually needs. I listed every voice the assistant could plausibly have, 44 candidate descriptors, cut the list to twenty against , grouped those by the job each word does, and weighted one winner per group. I removed every descriptor implying expertise Kedara does not hold, and every descriptor implying detachment.
Four descriptors survived: helpful, empathetic, direct, concise. The follows from them: open by acknowledging the situation and asking about it, deliver the substance without padding, close with one clear next action.
Why empathetic and direct together?
Caregivers arrive stressed, and flat clinical text reads as indifference. They also arrive needing to do something, so agreeable filler spends attention they do not have. show why each one fails. I excluded flattery, praise for asking, exclamation marks, reassurance beyond what is true, and any claim of expertise the product does not hold.
I designed each answer to resolve into a component, and gave each component a rule for when it applies: a step checklist for anything done in order, a comparison view for choosing between options, a do's and don'ts card for preventing mistakes, a resource card for action outside the chat, and plain text for acknowledgement.
Why question type decides the format.Structure produced by a model varies between answers to the same question. Structure assigned by question type does not vary, so a caregiver gets the same treatment every time she asks the same kind of thing. The dosage rule is the proof: in ChatGPT it hides in a paragraph; in Kedara it sits in a don'ts card where a skimming reader still passes it.
Type and contrast. I set body text at . The smallest text, the 14px pill labels, still . Testing kept returning to eye strain, so I made one idea per block a rule, not a preference. Three suggested prompts appear before the first message, because caregivers often arrive knowing they are overwhelmed without knowing what to type.
I drew the line between informing and prescribing. I designed the answers to explain what a medication schedule involves without setting a dose, and to compare two facilities without choosing one. The limit is a design decision I made early, not a disclaimer added late.
I put the citation beside the claim it supports. I designed the citation pill and the resource cards you see beside this text so that a caregiver who wants to check a source can, and one who does not still sees that a source exists. That is my answer to the research finding that this group trusts answers without knowing where they came from.
Some questions must not get the most useful-sounding answer. I designed the boundaries for those.
Self-harm is handled with maximum sensitivity. When a conversation turns toward self-harm, the assistant does not advise. I designed it to acknowledge, stay warm, and nudge strongly toward , which I made a first-class resource category, not a buried link.
Emergencies get detected and handed off. When a message reads like an emergency, the assistant urges the caregiver to call for help, and not only 911: it surfaces the specialized services that fit the situation better than a general emergency call.
This is the trade I accepted. In these conversations the safest answer is deliberately the less useful-sounding one. Kedara stops being a guide and becomes a bridge to a human. A confident answer here would be the most dangerous thing the product could produce.
A caregiver asks how to look after her father at home. The persona acknowledges the situation and asks what he needs help with. The question type routes the answer into a checklist with a do's and don'ts card. The dosage rule sits in the don'ts. Citation pills sit beside the sourced claims. The closing line names one next action, the scope note names the doctor, and if the conversation ever turns into a crisis, the product stops guiding and hands off.
Each layer does a different job. The persona gets the answer read. The structure gets it understood. The sourcing makes it checkable. The guardrails decide when Kedara must stop answering. Remove one and the others break.
The demo beside you is that full answer. Switch a layer off and watch it break.
Shipped across desktop and mobile web; the mobile app is scheduled for Q3 2026. My time on it: January to May 2026, AI Product Designer, Voice AI.
The honest limit. The guardrails and answer content have not been through formal clinical or Medicaid review. That review was being arranged when I left. My design work structures the answer, sources it, and marks its boundaries. It does not make the underlying content clinically verified, and the product does not claim it does.