The people who raised us deserve to grow older on their own terms — at home, safe, and never alone with a worry at 2am. Ardia is building a gentle, always-there layer of intelligence that helps older adults stay safe, independent and connected, and quietly brings family and clinicians in before a small change becomes a crisis.
A non-diagnostic elder-care layer that pairs a guardrailed voice companion with a trained movement classifier to surface safety flags and caregiver alerts — decision support, never a diagnosis.
A Claude/Gemini persona under guardrails runs daily check-ins; the LLM never decides safety. Each turn runs scripted screens — stroke FAST, suicidality via the C-SSRS screener — through a deterministic router that escalates to 911 and the caregiver. Mood prompts feed PHQ-9/GAD-7; cognition concerns route to clinician-administered MoCA. No published accuracy.
A scikit-learn classifier on public UCI HAR (phone accelerometer/gyroscope) recognises activity type — measured 95.45% on a subject-independent held-out split of volunteers aged 19–48. That accuracy is activity recognition, not gait, fall, or deterioration detection, and not elders; those are modelled targets pending elder-cohort retraining under BAA.
Each module rests on named authorities, not invented logic: AGS Beers Criteria for medication-safety flags; PHQ-9/GAD-7 (mood), MoCA (cognition), Braden Scale (pressure-injury risk); mapped to CMS care-management codes — CCM 99490/99439, RPM 99453/99454/99457/99458, TCM 99495/99496, AWV G0438/G0439, ACP 99497/99498.
Findings surface to care teams and caregivers over HL7 FHIR (US Core), with SDOH captured as ICD-10 Z-codes (Z55–Z65) to route social-need referrals. Every red-flag event and PHI access writes an audit log; escalations fan out to 911 and the named caregiver. Anything predictive or device-like is FDA 510(k)-gated and deferred, not shipped.
Guardrailed voice companion for check-ins; hands every safety call to a deterministic red-flag router, never escalating itself.
Recognises activity type on UCI HAR (95.45%, subject-independent, ages 19–48); gait/fall/deterioration on elders unproven.
Rules-based spine: FAST/C-SSRS triggers escalate to 911 and caregiver; the LLM never controls emergency escalation.
Train and test Cadence on public UCI HAR; wire Aria to a deterministic FAST/C-SSRS → 911/caregiver spine; map each module to its CMS codes and guidelines. Non-diagnostic, public data only, no PHI.
Under signed BAAs, retrain Cadence on a consented elder cohort and measure elder-specific accuracy; connect care-team/caregiver workflows over FHIR; supervised pilots with clinician oversight on every flag.
Broaden EHR integrations and module coverage; formally evaluate an FDA 510(k) pathway before any predictive or device-like deterioration alerting ships. Everything remains decision-support until cleared.
Aria is the warm voice an older adult actually talks to — a companion first, and a safety net always. She checks in gently each day, helps with medications and reminders, and listens for the handful of red flags where minutes matter. Aria is a persona model running on a frontier LLM (Claude Opus/Sonnet or Gemini Flash/Pro) behind hard guardrails; she has no published accuracy yet and never makes a diagnosis. What she reliably does is notice, reassure, and escalate.
A short, human conversation — “How did you sleep? Any pain today?” Aria remembers yesterday, so a change stands out instead of getting lost.
Warm nudges for doses, appointments and hydration — and a flag to the caregiver if a critical medication is repeatedly missed. Support, not surveillance.
When something sounds wrong, Aria runs a structured, non-diagnostic check — stroke FAST signs, and the C-SSRS for suicide risk — to decide how urgently a human is needed.
If it's a genuine emergency, Aria does the single most important thing immediately — escalates to 911. No diagnosis, no delay, no judgement call left to the older adult alone.
Family or the care team is looped in with what happened and what to do next — so nobody finds out too late, and nobody has to hover all day to feel safe.
Aria's daily contact, movement signals and follow-ups are built to become the documented touchpoints these care-coordination programs require — so the care team gets a clearer picture of the patient between visits, and the reassurance reaches the family in the same loop. These are modelled care-pathway targets, not billed services — Ardia has no patients today.
No app to learn, no small buttons, no password. Aria meets an 82-year-old where they are — in conversation. The interface disappears, and what's left is someone who checks in, remembers, and knows when to call for help.
Functional decline in older adults is rarely a single dramatic event — it's a slow, quiet drop in how much someone moves through their own day. Cadence is built to sense that drop early, from nothing more than a phone carried in a pocket — so a care team can act while the change is still small and reversible, rather than after a fall or a hospitalisation. It is not a fall detector; it watches the trend, not the moment.
Unlike the persona models, Cadence isn't an LLM — it's a scikit-learn classifier trained from scratch and scored on a public benchmark. On the UCI HAR dataset, evaluated subject-independently (the test people never appear in training), it reaches 95.45% held-out accuracy — macro-F1 0.9545, a 99.63% Active/Sedentary/Resting rollup, and an 85.71% worst-subject floor. That's a measured fact, reproducible from public data.
See the full method, confusion matrix and test suite on the Test Results page →
UCI HAR's volunteers are aged 19–48. Older adults move differently — slower, with different gait and posture — so this 95.45% proves the method and pipeline, not that Cadence is 95.45% accurate on a frail 82-year-old, and we won't imply it is.
Before any elder deployment, Cadence will be re-trained and re-validated on older-adult movement data under a BAA, on this same subject-independent protocol. Until that's measured, elder-population accuracy is an unproven target. Cadence is non-diagnostic and not a fall detector.
Aria and Cadence are the companion and the movement layer; around them sits a clinical scaffold of nine focus areas, each mapped to instruments and programs clinicians already know. Every module is a modelled care pathway — a design target, not a shipped or validated product.
Flags risky and unnecessary drugs using the AGS Beers Criteria.
Documented touchpoints across CCM · RPM · TCM care programs.
Annual Wellness Visit and Advance Care Planning — AWV · ACP.
Depression, anxiety and cognition — PHQ-9, GAD-7, MoCA — plus caregiver support.
The everyday risks that compound quietly — falls, nutrition, sleep.
Long-term condition support — heart failure, AFib, diabetes.
Pressure-injury and continence risk via the Braden Scale.
The senses that keep a home safe — vision, hearing, home safety.
Reaching everyone — SDOH, telehealth, palliative support.
The number of older adults is rising faster than the number of people able to care for them. The answer can't be more alarms and more dashboards — it has to be something warmer: a system that pays quiet, patient attention, so that independence lasts longer and help arrives sooner. That's the entire point of Ardia's aging work — technology that lets a family worry a little less, and a parent stay themselves a little longer.
We're pre-revenue and seed-stage in Dallas–Fort Worth, building this honestly and in the open — every claim sourced or marked as a target, every emergency routed to 911, nothing dressed up as a result it isn't. If you're a clinician, a health system, or a partner who can help validate Cadence on older-adult data under a BAA, we want to talk.
