BMT-01.02 Executive Summary#
BlueMirror.tech | May 2026#
Margaret’s cardiologist sees her once a year for thirty minutes. He has nine hundred patients. The arithmetic is unforgiving: his attention to Margaret across a year totals thirty minutes, and the remaining 525,570 minutes belong to no one. The minutes when her blood pressure climbs three points across four readings and no one notices. The minutes when her weight gain begins three weeks before the ER visit and no one connects the two. The minutes when her medication timing slides from 7:00 to 7:45 to 8:30 because no one is watching. The cardiologist decides. The concierge watches.
The health concierge is the most complex single agent in the BlueMirror system: six infrastructure agents, five small language models, a mixed autonomy profile that runs high for routine monitoring and low for care transitions. The complexity maps to the complexity of the gap. From Margaret’s perspective the agent is one entity, but six infrastructure agents do the work beneath. The Medication Manager owns reminders, refills, adherence tracking, and interaction checking, running at 0.75 autonomy on the edge. The Symptom Monitor tracks reported symptoms and looks for pattern signatures like the five-day fatigue trend that often precedes infection in elderly diabetics. The Vital Signs Analyst ingests blood pressure, glucose, weight, and heart rate from connected devices and flags deviations against the person’s own baseline. The Exercise Monitor assesses mobility patterns including gait variability and transfer time from sit to stand. The Appointment Coordinator manages scheduling and pre-visit preparation. The Care Transition Manager handles discharge planning at 0.25 autonomy in the cloud, where every meaningful action requires human approval.
The SLM stack is sized as deliberately as the agent decomposition. The Medication Advisor at 150M parameters is the smallest size at which a model reliably handles polypharmacy reasoning for a population whose median patient takes seven concurrent medications. The Cognitive State Estimator at 200M is shared with the cognitive concierge so that one assessment drives behavioral adaptation across thirteen agents. The Safety Filter at 100M parameters runs with a non-negotiable 25ms latency target because it sits in the response path on every interaction. The Intent Classifier at 150M parameters routes requests to the right capability. The Response Generator at 400M is the largest because the prose quality at the surface determines whether Margaret keeps engaging. Total footprint is one billion parameters across five models, with cumulative inference under 325 milliseconds on a current-generation tablet. Edge deployment is what makes the privacy guarantees credible: the medication context never leaves the device.
The autonomy gradient follows risk, not convenience. Medication reminders execute autonomously. Refill requests execute autonomously with notification. Symptom pattern alerts to family follow a 24-hour delay protocol that allows the person to address the underlying condition or update the agent before family gets pulled in. Care transition planning requires human approval at every step, because care transitions are where elderly patients are most often harmed by mistakes and the consequences of action without approval are too severe to accept.
The clinician interface defines three boundaries. The agent reads the active medication list, the problem list, the allergy list, recent labs, and post-visit notes through FHIR R4, scoped to providers the patient has explicitly added. The agent writes adherence data, symptom reports, and structured pre-visit summaries back to providers who accept FHIR write-back, a partial reality at major academic centers in mid-2026 and a future state for most community practices. The agent never diagnoses, prescribes, or contradicts a clinical order. Crossing that line triggers FDA medical device classification and changes the architecture from a software product into a regulated device subject to 510(k) clearance. The product roadmap holds the line. When Margaret asks “Should I increase my blood pressure medication?” the agent does not answer the question as posed; it surfaces the trend, prepares a structured question for her cardiologist, and offers to send it through the patient portal. The decision is the cardiologist’s. The preparation is the agent’s.
The cognitive capacity overlay runs continuously. When the Cognitive State Estimator reports a low-capacity day, reminders use shorter sentences, questions become more concrete, the visual layout shifts toward larger fonts and fewer options. The agent does not announce the change. A system that requires the person to switch modes is a system that fails when the person’s capacity to switch modes itself declines.
The honest limitation: the Cognitive State Estimator is not a clinical diagnostic. It cannot distinguish between mild cognitive impairment, a low-sleep day, depression, and an acute infection. When it detects significant deviation from baseline, it does not diagnose; it surfaces a question to the clinical care team and adjusts behavior in the meantime. The clinician decides. The agent watches. The architecture is honest about what depends on what: a health concierge running without sensor data is a different system, still useful but not the same.
For the full treatment of the SLM stack, the autonomy gradient, the FHIR boundary, and the cognitive capacity overlay, read the complete article on BlueMirror.tech.