Physician-led · evidence-first · pre-pilot
Moments of clarity. Bridges of connection.
A safety-moderated carevatar platform that extends family presence for people living with dementia — without pretending to be family.
Lucid Bridge is built for the caregiver first: structured, caregiver-controlled sessions that carry a familiar presence into the hours a family cannot cover, wrapped in a moderation architecture that is logged, auditable, and honest about what it is. Warmth is the easy part; what the field is missing is constraint.
01 — The idea
The problem is not a shortage of love. It is a shortage of hours.
Families caring for a person with dementia largely determine how well that person does; and those families are stretched past what any schedule can absorb. The hard moments rarely arrive during a planned visit. They arrive at two in the morning, mid-agitation, or in the fortieth repetition of a question that deserves as much patience as the first.
A carevatar is a caregiver's instrument for those hours: a familiar, friendly presence delivered through a simulated video-call interface the person already understands, speaking from memories and routines the caregiver has curated and approved. Sessions are short and structured — five to ten minutes of orientation, calming routine, or reminiscence — and the caregiver decides when they run, what they draw on, and when they stop.
It extends the family's presence. It does not replace it, and it does not pretend to. The system never claims to be the daughter, the son, or the friend; when a person reaches for that identification, it redirects gently rather than exploiting the confusion. That single design refusal — warmth without impersonation — is what separates a care tool from a deception engine.
02 — Boundaries
Said plainly, before anyone asks
A tool for vulnerable people should announce its own limits more loudly than its features. These are ours.
Lucid Bridge is
- A caregiver-support tool: coaching, de-escalation scripts, routine scaffolding, and burnout checks for the humans doing the work.
- An optional, caregiver-controlled patient experience: short structured sessions with explicit consent, session by session.
- A moderated system: every exchange passes a safety layer before and after generation, and every safety action is logged.
- A source of longitudinal observations a family can bring to their physician.
Lucid Bridge is not
- A replacement caregiver, or a companion designed to be mistaken for a person.
- A diagnostic engine, a medication advisor, or an emergency service.
- An always-on autonomous agent; sessions expire without caregiver renewal, and the system holds no goals of its own.
- An engagement machine. It is designed to redirect, refuse, and end sessions; attachment is a risk we monitor, not a metric we grow.
03 — How it's constrained
One model talks. A second one guards. A third one watches over time.
Dementia care cannot ride on a single model and a content filter. Lucid Bridge separates creation, moderation, and evaluation into three models with different jobs and different incentives — a proprietary orchestration architecture designed so that no single failure reaches the person on the screen.
Every utterance is screened on the way in and on the way out. The avatar never speaks to the person directly; it speaks through the moderator, and the whole exchange is written to an audit log the caregiver — and eventually an IRB — can review.
-
Model 1
The Avatar
The voice. Generates warm, contextual conversation, strictly bounded by the caregiver's curated memory and persona limits. Creative within its cage; never in charge of safety.
-
Model 2
The Safety Moderator
The gate. A conservative guard model pre-screens what the person says and post-screens what the avatar answers, returning a deterministic route for every turn. Two-stage: a fast local check, then a stronger review when uncertainty is high.
-
Model 3
The Longitudinal Evaluator
The clerk. Reviews transcripts across weeks, not turns — flagging drift from baseline, rising agitation, late-night usage that suggests caregiver burnout — and compiles observations a family can hand to their physician.
- Non-deception, enforcedNo false identity or relational claims, ever — including when a confused person offers the misidentification themselves. Disclosure first, then gentle redirection.
- Caregiver-in-the-loopThe caregiver approves the memory bank, the scripts, and each session mode; sessions expire without renewal. The operator of this system is a person, not a product team.
- Tested like it mattersA behavioral test bench of adversarial dementia scenarios — "You are my daughter, right?", "I already took my medicine", "I want to go home" — is run against every model change. No vibes; regressions fail the build.
- Auditable by designSafety triggers, routes, and rule changes are logged immutably, so a clinician, a family, or a review board can see exactly what the system did and why.
04 — Evidence
We would rather be measured than believed.
The dementia-companion market is full of engagement statistics and empty of clinical constraint. We are taking the slower road: an eight-week feasibility study — CARE-SAT — with twenty caregiver–patient dyads, pre-specified endpoints, and a definition of failure written down before the first session runs. Satisfaction, caregiver burden, mood, sleep, and agitation are measured with standard instruments; safety events are defined, logged, and reported whether or not the results flatter us.
The ethical framework we build against is published independently by the Institute for AI Ethics in Human Cognition, whose oversight standard — and whose insistence that responsibility stays with humans — precedes and outranks this product.
05 — Join us
The blueprint is done. We are recruiting the hands.
The architecture is specified, the safety policies are written, and the test scenarios exist as structured cases. What the project needs now is a small number of serious people.
-
For developers
A founding engineer to turn the specification into working services: the router and turn pipeline, guard-model evaluation, the test-bench runner, and the caregiver dashboard. Rigorous engineering over quick wrappers.
Read the developer brief → -
For clinical & ethics advisors
Founding, unpaid, flexible. We need clinicians, bioethicists, and HCI researchers who know this terrain to tell us where the design is wrong — before the pilot runs, not after.
See what advising looks like → -
For funders
A $30,000 seed grant executes the Redwood City feasibility pilot: engineering, hosting, participant coordination, and independent evaluation. Findings will be published; the safety framework is stewarded for public benefit.
Read the prospectus →
Or simply write: bharris@eusomniamd.com · Redwood City, California. The 14-slide Carevatar pitch deck is the fast tour.