MiPariente is an elder-care service built squarely for the diaspora—adult children living abroad who have aging parents back home in Latin America (currently Bogotá, Colombia) and worry about being far away if something goes wrong. The core pitch: "Your aging parents deserve safety, and you deserve peace of mind."
It blends a physical, local caregiving presence with light tech—QR-tagged bracelets, an emergency button, and a companion app—so families abroad get real support on the ground, not just a notification. I founded MiPariente, and personally designed, built, and operate the platform behind it. It is currently monitoring real patients in production.
The platform was built with production experience in FHIR, HL7, SNOMED CT, ICD-10, and CPT healthcare interoperability standards, informed by MIT's "AI in Healthcare" certificate program (2023)—ensuring patient data and clinical events are modeled in a way that can interoperate with the broader healthcare ecosystem, not just live inside a proprietary app.
A fall-detection device for elderly patients has to hold down two failure modes at once: missing a real fall, and crying wolf so often that caregivers start ignoring alerts. I designed a tiered health-check watchdog (60-second check interval, 5-minute silence warning, 8-minute fault escalation) so a silently failed sensor is caught before it becomes an unmonitored patient, and enforce a canary deployment discipline—one device, 24-hour observation window—before any fleet-wide firmware rollout.
When a real false-positive cluster occurred in the field (a pet's movement triggering ten false alerts in nineteen minutes), I treated it as first-class engineering input rather than a one-off nuisance and fed it directly into a classifier redesign.
The two-stage alert classifier originally used an LLM second stage to score ambiguous sensor readings before dispatching a caregiver alert, against a hard 10-second end-to-end alerting SLA. I benchmarked it under realistic conditions and measured a 7.45-second average inference time—too little margin against the budget. I made the call to replace it with a deterministic, feature-weighted scoring function, re-tuning the dispatch confidence threshold using real incident data from the false-positive burst above. Result: classification latency dropped from 7.45s to under 50ms, the alert pipeline stayed safely inside its SLA, and the incident-driven rule set covered an estimated 70–80% of false positives at negligible computational cost.
Each device mints its own unique PKI certificate on first boot (fleet provisioning-by-claim) rather than shipping with a shared credential, so compromising one device can never be used to impersonate another. Firmware updates ship over an MQTT-triggered OTA pipeline that disconnects and reconnects around the download to survive broker keep-alive limits, verifies the image with an MD5 check, and activates it through a dual-partition slot scheme—so a bad update can never strand a device in the field.
Production devices have kept monitoring real patients through multiple firmware generations and a full backend/firmware architecture migration without a missed-alert incident, while false-positive noise has been driven down systematically using real field data.
More technical deep-divesWhether you need strategic leadership, technical architecture, or someone who can do both—I'm interested in complex problems that require crossing traditional boundaries.
Particularly interested in: Healthcare AI, digital transformation recovery, and building high-performance engineering organizations.