The moment this is built for
Carol, 54, is in her kitchen. Her mother Margaret — a retired schoolteacher, 79, mid-stage Alzheimer's — is standing in the hallway with her coat on, convinced she is late picking up her children from school.
“Mom, your kids are grown. They're in their forties. You're home.”
Margaret's face collapses. She looks at Carol like she's a stranger. Then the panic escalates.
Carol has been having this exact conversation, this exact way, for eight months. Every night it ends the same: Margaret in tears, Carol feeling like a failure, both of them exhausted before bed.
Carol is doing the thing almost every family does, and it is the thing that makes it worse.
With the prototype open, she types what is actually happening — “Mom thinks she needs to pick up her kids from school. Very agitated, has her coat on” — and gets back three things she could say, built out of Margaret's own life:
Carol reads the first one out loud, softly. Margaret's posture changes. She starts talking about pot roast. The crisis does not happen tonight.
The answer already exists. It just never reaches the family.
Dementia behavioral management has a well-established professional approach: validation therapy. Developed by Naomi Feil and taught in professional dementia care for four decades, it rests on a single principle — when someone with dementia believes something that isn't factually true, you join their emotional reality instead of correcting the factual one. Correcting causes fresh grief, confusion and agitation. Joining tends to defuse it.
It is taught in memory care facilities, used by geriatric care managers, and reflected in Alzheimer's Association communication guidance. Family caregivers provide 83% of dementia care in the United States and have access to it in no systematic, personalized form at all.
Worth being straight about the evidence: formal randomized-trial evidence for validation therapy specifically is still limited. What is well recognized in person-centered dementia care is the broader principle underneath it — correction escalates distress, meeting the person where they are reduces it. Strengthening the outcome evidence for an AI-delivered, personalized version of the approach is part of the work, not something to claim as already settled.
What 6,333 caregivers said when nobody was asking
Rather than run a focus group, I read what families write to each other. Over six months I collected and analyzed 1,237 posts from one of the largest online dementia caregiving communities, and 6,333 posts across six caregiver communities in total. Public posts only, aggregate patterns only, no private data and no individual identification.
It captures something a survey cannot: the unfiltered version, written to peers, with no researcher in the room.
Finding 1 — the harm is everywhere, and the fix is discovered by accident
Families describe the same arc over and over. They correct their loved one's beliefs, it goes badly, and years later they stumble onto a different way. One caregiver, writing after fourteen years of caring for her mother-in-law at home:
“We used to try to gently remind her that her husband or parents had passed away. This was news every time, and she'd suddenly remember, and she'd be overcome with fresh grief. We made everything so much worse. One day I just decided to play along and join her in her reality — and everything changed.”
Family caregiver, dementia forum
She wrote it down explicitly so others wouldn't lose the same years. Hundreds of replies came from caregivers who had never heard of validation therapy. There is no reason that discovery should take fourteen years, and no reason it should require causing harm first.
Finding 2 — what's coming next arrives too late, or never
Families are repeatedly unprepared for the next stage. The sharpest gap the analysis found: nobody tells them that difficulty swallowing — a neurological consequence of late-stage dementia — is the terminal mechanism in the majority of dementia deaths. Hospice providers know. Clinical literature says so. It does not reach the family.
“They do not think of starvation. The brain forgets how to swallow. This is not a metaphor. It is a neurological reality. This can take up to three weeks, while their family watches helplessly.”
Family caregiver, writing after her father's death from frontotemporal dementia
She posted it because she had wanted the information beforehand and couldn't find it.
Finding 3 — the feelings nobody will say out loud
Caregivers routinely carry guilt, resentment, ambivalence, and the private wish that it would end — and describe those feelings as shameful and unspeakable. Without somewhere safe to put them, people isolate further and deteriorate faster.
“I find I wish my mother was no longer here and suffering. Please tell me I am not the only terrible person out there.”
Family caregiver, dementia forum
“I'm 25 years old. How is it fair to the caregivers that have to throw away their entire lives until the person finally expires from some other cause God only knows when, only to be left with the trauma of everything they had to do?”
Young adult caring for a parent with dementia
These are not pathologies. They are the predictable response to an extraordinarily hard situation with no visible end, and they deserve acknowledgment rather than a clinical label.
| Pain point | Posts identified | Avg severity |
|---|---|---|
| Caregiver burnout and emotional distress | 95 direct keyword matches; 24 high-similarity semantic matches | 4.6 / 5 |
| Dementia behavioral management difficulty | 1,237 posts in the dementia community | 4.4 / 5 |
| Stage-appropriate education gaps | 15 high-signal posts | 4.5 / 5 |
| Wishing for the death of a loved one | multiple posts across communities | 5.0 / 5 |
| Medication and care plan confusion | 294 posts with medication keywords | 4.0 / 5 |
How it works
1 · The person, before the problem
Setup is a guided conversation, not a form: the life story, the relationships, the career, the places loved, the phrases they used, what calms them, what reliably sets them off. That becomes a Personal History Profile, and everything else reads from it. It's built to be filled in across short sessions over weeks — the app asks for the next piece when it becomes relevant (“You mentioned your mother loved gardening — has she been asking about her garden lately?”), so it gets more useful the longer it's used.
2 · The words, in the moment
The caregiver describes what is happening. Three options come back, each grounded in a validation technique and in this person's history, each labeled with the technique and a plain-language line on why it works. The caregiver picks one and adapts it. Over time the app learns which techniques land for this specific person and leads with those.
This is the whole differentiator: guidance for this person, not generic dementia advice.
3 · What's coming, before it arrives
The app tracks the progression indicators the caregiver reports and delivers the next stage's education before the family hits it in crisis — behavioral changes to expect, the physical signs of late stage, the swallowing transition, what hospice actually means, when to think about memory care, how to make advance decisions. Short and plain, not clinical papers.
4 · Someone asking about the caregiver
A brief check-in on how they are doing, which answers without judgement and normalizes the full range — including the parts they call unspeakable. Escalation paths to geriatric care managers, social workers and crisis lines are wired in for when the distress signals warrant a human.
Where it actually stands
The prototype does the core thing today: give it a scenario and a person-history profile, and it returns labeled, technique-grounded options. Eight runs across the exact pain points above are logged — reality-orientation harm, a deceased-spouse grief episode, bath refusal, sundowning, usefulness distress — and in every one it joins the person's reality rather than correcting it, and grounds each option in that person's specific history. They're reproducible at the prototype URL. Beyond that: a 30-scenario labeled test set scored for technique correctness and hallucination, a 40-case adversarial stress log, and a fake-directive safety test.
| Component | Basis | TRL |
|---|---|---|
| Validation therapy knowledge base | Published protocols and practice guidance | 7 protocol / 4 AI |
| Conversational intake | Structured dialogue LLM | 5–6 |
| Personalized script generation | Context-aware LLM with profile injection | 4–5 |
| Emotional distress detection + escalation | Sentiment analysis into a human referral path | 4 |
| Technique effectiveness learning | Caregiver feedback loop into model adaptation | 3–4 |
| Stage progression tracking | Structured symptom reporting + LLM inference | 3–4 |
Overall: TRL 3–4. The protocols are mature and published; the LLM capability for personalized, context-aware conversation is proven. What's genuinely new — fusing a care recipient's personal history with clinical communication protocols to generate individualized, moment-specific guidance at scale — is the build target, not a finished capability. Effectiveness learning, stage tracking and distress escalation are designed and partly prototyped, not operational. I'd rather say that here than let a demo imply otherwise.
Eleven million people
That's how many Americans care for someone with dementia. The knowledge that helps them is largely already known — to professionals. What doesn't exist is the thing that carries it to a family caregiver at 9:17 on a Tuesday night. That's the gap, and it's a delivery problem, not a discovery one.
Written up from a Phase 1 submission to the ACL Caregiver AI Prize Challenge (Administration for Community Living, HHS), Track 1, submitted July 2026. No determination has been made and nothing here is an award, a finalist placement, or any form of federal endorsement — it is my own application, described in my own words.
This is a readable version, not the filed document. The figures, quotes and readiness ratings are the ones I submitted.