Proche Aidant is a conversational AI assistant purpose-built for people in the early stages of dementia and their caregivers. Unlike general-purpose assistants, it remembers who you are, tracks your medications, and responds when you're in distress.
This isn't a mockup or a prototype. Proche Aidant is deployed on Google App Engine and available right now. Say your name, ask about the weather, log a medication — it remembers everything.
A working paper on the system's hybrid NLU architecture and usability evaluation has been submitted to HHAI 2026 — an international peer-reviewed conference on human-AI collaboration, hosted in Brussels.
No login, no forms. Just say your name and Proche Aidant remembers. Your profile builds naturally through conversation.
Medications, schedules, emergency contacts, entertainment preferences — all stored persistently and recalled across sessions.
Proactive medication reminders, distress detection that routes to your personal emergency contact, and a companion that's always there.
Say "my name is Priya" on Monday — Proche Aidant greets you as Priya on Thursday. Names, medications, contacts, and preferences persist indefinitely.
Keyword and semantic pattern matching detects fear, confusion, or emergency signals and routes directly to the user's stored emergency contact.
"I took the white pill" is understood as a medication log. Proactive push reminders via Server-Sent Events ensure doses aren't missed.
Designed for users who can't produce precise commands. Indirect, vague, and conversational inputs are handled gracefully by the NLU routing engine.
Movie recommendations with genre, language, and mood filters. Real-time weather queries. Daily companionship, not just clinical utility.
Distress detection, toxic content filtering, and graceful degradation — if any service fails, the system returns a calm, reassuring message.
| Feature | General Assistants | Proche Aidant |
|---|---|---|
| Memory | Resets between sessions. Doesn't remember your name. | Persistent across all sessions. Remembers name, meds, contacts. |
| Input tolerance | Requires precise commands. Fails on vague input. | Understands indirect phrasing. "I took the white pill" works. |
| Safety | No distress detection. Routes to generic helplines. | Detects distress and routes to personal emergency contacts. |
| Designed for | General population. Assumes working memory. | Early-stage dementia. Accounts for cognitive decline. |
| Model | Large, resource-heavy (GPT-4 scale). | Llama 3.1 8B — small, fast, sustainable by design. |
Caring for a parent or loved one with dementia. Proche Aidant reduces the daily burden.
A companion that preserves independence — no technical literacy required.
A deployed proof-of-concept for conversational AI in cognitive care, with published evaluation data.
Proche Aidant is a deployed, live system — not a prototype. Here's the technical detail for those who want to go deeper.
| Participant | SUS Score | Band | Profile |
|---|---|---|---|
| P1 | 95 | Excellent | Adult, caregiving background |
| P2 | 87 | Excellent | Biotechnologist, patient care experience |
| P3 | 85 | Excellent | Elderly participant, aged 67 |
| P4 | 85 | Excellent | Cares for elderly family member |
| P5 | 83 | Excellent | Cares for elderly family member |
| Mean | 87.0 | Excellent | SD = 4.69 (highly consistent) |
Replace text input entirely for late-stage patients. NLU models for disordered speech are an open research problem.
Learn a user's idiosyncratic language patterns over time. Personalised language modelling for cognitive impairment.
Smart medication dispensers, door sensors, wearable fall detectors — continuous passive monitoring.
Structured study with actual dementia patients under ethics approval — transforming proof-of-concept into evidence.
Whether you're a caregiver, researcher, or potential partner — we'd love to hear from you.
✉ Contact Us