Carevisor — Reducing re-hospitalization by driving adherence with a protocol-bound AI caretaker at B2C scale

Overview

Carevisor helped elderly, co-morbid, high-incidence patients reduce re-hospitalization by improving medical adherence through an “at-home” medically trained AI caretaker: a video avatar that proactively called patients, triaged issues, and escalated to human clinicians when thresholds were met.

The operational north star was adherence (meds, biometrics, activity, appointments, family support). The revenue/outcome north star was lower re-hospitalization, with adherence as the leading indicator we could move day-to-day.

The product ran as a HIPAA-compliant, B2C-scale system designed for ~10M concurrent users, integrating IoT telemetry, trend detection, patient/caregiver apps, and a protocol management backend for provider/physician customization.

Situation & Stakes

Target users were elderly patients with multiple conditions: high risk, low tolerance for UX friction, and limited comfort with smartphones.
Success required behavior change, not information: the system needed to earn engagement and sustain it long enough to affect adherence.
Clinical liability was the core constraint: any automated action had to be attributable to provider-defined protocols, not “AI judgment.”
Escalation performance was non-negotiable: < 3s SLA from triage completion to placing a human-support call.
HIPAA compliance was table stakes, but not sufficient; auditability and determinism were required for trust and contracting.
The platform had to support deep customization (provider → physician → patient) without exploding UX complexity or operational burden.

Observations & Decisions

We optimized the primary UX for “talk and listen,” treating UI feeds as a fallback and learning surface, not the main interaction model.
Early designs leaned toward smartphone-native “feeds + cards” to manage status and touchpoints. I pushed back: that pattern selects for patients already comfortable navigating apps—exactly the cohort least in need of an avatar caretaker. For our common-denominator user, the reliable interaction primitive was conversational audio/video. We made the avatar-led call flow the default and demoted feeds to: (1) error recovery, (2) summary and reassurance, and (3) a surface to reinforce habits and close loops back into the caretaker. In studies, this materially improved initial engagement and sustained use.

We treated video as an adoption wedge, not a permanent requirement, because early emotional comfort was a leading indicator for long-term adherence.
Text was supported as fallback, but was a poor fit for many elderly users. Audio worked, yet video reliably increased early trust and willingness to participate—patients felt “cared for,” not “prompted.” A pattern emerged: many users disabled video after weeks, but disabling video from day one reduced engagement. We designed the system so video could gracefully step down over time without breaking the conversational experience.

We made the system operationally deterministic by binding every action to a provider-defined protocol, eliminating agentic judgment in clinical decisions.
Given the risk profile, the system could not “decide like a clinician.” Thresholds and escalation rules were fully configurable by provider/physician, with sensible default profiles by disease pattern. The AI executed protocols and triage logic deterministically; liability remained anchored to explicitly defined and contractually understood protocols rather than emergent behavior.

We required every automated action to be explainable back to protocol—and we instrumented verification loops to detect and correct drift.
For a healthcare agent to be shippable, “what it did” wasn’t enough; it needed to show “why it did it” in terms clinicians could accept. We implemented an action narrative that referenced the triggering protocol elements, then validated these explanations statistically across production-like scenarios. When violations appeared, we treated them as correctness bugs: isolate, reproduce, and correct behavior. Where judgment was unavoidable in ambiguous inputs, we used ensemble-style redundancy to reduce brittle failure modes (at the cost of complexity and latency budgets).

We designed escalation as a first-class, performance-budgeted subsystem, not a feature, because false negatives carry asymmetric risk.
In adherence workflows, false positives waste human time; false negatives can lead to harm. We biased toward escalation when uncertainty crossed thresholds defined in protocol, and we engineered the handoff to meet the sub-3s call placement SLA after triage completion. This drove architecture choices: tight instrumentation, predictable execution paths, and clear boundaries between automated triage and human responsibility.

System Design Overview

The platform was a closed-loop care orchestration system with explicit boundaries between:

Patient interaction layer: proactive calls via video avatar (audio fallback), plus patient/caregiver mobile apps for summaries, reminders, and recovery paths.
Telemetry & signals: IoT for continuous biometrics, plus activity and appointment adherence signals.
Trend detection: ML-driven trend and anomaly detection used to trigger protocol paths and escalate earlier when patterns changed.
Protocol engine: hierarchical configuration (provider → physician → patient) with disease-profile defaults. Protocols defined thresholds, escalation criteria, and allowed actions.
Escalation & clinician workflow: deterministic routing into the human support hierarchy, built to meet strict latency SLAs and preserve audit trails.
Governance & compliance: HIPAA controls plus auditability: action logs, protocol references, and traceable handoffs.

Key control concept: the “AI” was treated as a capability inside a system with explicit guardrails—protocol definitions and deterministic execution were the safety boundary.

Impact & Outcomes

Reduced re-hospitalization by driving measurable gains in medical adherence
Achieved B2C-scale (~10M concurrent users) while maintaining HIPAA compliance
Registered one of the world’s largest providers
Deployed during COVID with universities and government programs
Company was acquired; resulted in an 8-figure exit.

Reflection

What worked:

Treating elderly UX as its own domain with its own primitives—conversation first, UI second—created adoption where “standard app UX” failed.
Making clinical behavior protocol-bound aligned sales to market needs for contractual liability containment
Designing escalation as a performance-specified subsystem avoided the common trap of “handoff as an afterthought”

What would fail if copied blindly:

Video avatars are not universally additive; here they were an adoption wedge because the target users needed emotional reassurance, not interface efficiency.
Protocol determinism only works if providers will actually own protocol definitions and accept operational responsibility; without that, liability and governance collapse.
Sub-3s escalation requirements force architectural discipline; teams that treat latency as “later optimization” will discover it as a safety failure.

Role & Scope

Product design lead: defined target-user UX primitives, interaction model, and engagement strategy.
Solutioning + architecture: shaped protocol-bound execution model, escalation boundaries, and compliance/audit requirements.
Head of engineering: led build across avatar interaction, apps, IoT integration, protocol backend, and operational readiness for scale.
Executive client interface + pre-sales: owned stakeholder alignment, provider requirements translation, and executive trust through delivery.