Neonatal Monitoring · Medical Device Regulatory · Clinical Evidence

The healthcare problems I care about,
and how I'm starting to solve them.

Agentic systems handle both — one unlocks the data, the other compresses the authoring. Lookup, cross-referencing, standards-pinned drafting — from the codebase, not someone's memory.

Chapter 01
Neonatal monitoring
SpO2 pipeline · Babysat
Chapter 02
Regulatory infrastructure
SaMD-OS · 5 specialist agents
Chapter 03
Clinical evidence design
Pathway Designer · FHIR CarePlan
Other work
Draft Day · QB Prospects
Nourish · flash_crash
5
Specialist reviewers
6
Standards referenced
21
Regulatory skills
3
Neonatal systems
01
Neonatal Monitoring

380,000 preterm infants go home each year with no personalized monitoring.

Providers discharge infants with BPD and CHD using fixed SpO2 thresholds that don't account for gestational age. A 90% alarm on a 34-week preemie fires constantly — creating alarm fatigue that masks real desaturation events. Parents get raw numbers with zero clinical context.

Insight
Baselines must be personalized. Preterm SpO2 ranges (92–95%) overlap with term "borderline" values (97–100%). GA-adjusted p5–p95 bands validated against Castillo 2008 and Dawson 2010 cut false alarms without missing true hypoxic events. Clinically significant desaturation requires both depth and duration — not just threshold crossing.
What I Built
SpO2 Pipeline: Rule engine → ML classifier → LLM-as-judge evaluation. HL7v2 interop, EHR-ready triage flags.

Babysat: React Native home-monitoring app. 14-day rolling baselines, AI narratives via Claude API, BISQ-R sleep assessment, provider handoff reports with tokenized viewer.
SpO2 Pipeline → Babysat →
02
Regulatory Infrastructure

SaMD teams can't ship because their regulatory artifacts are disconnected from their codebase.

Design controls live in spreadsheets. Risk files reference code that's already changed. Every 510(k) submission is a manual reconciliation exercise. Teams that shipped first have zero traceability between their code and the regulatory artifacts they need to produce.

Product Judgment
Built SaMD-OS — a regulatory operating system that generates IEC 62304 artifacts from the codebase and reviews them with specialist agents. Three reverse-engineering skills for teams that shipped before documentation: Code → SOUP Register, Code → Design Inputs, Code → Hazard Candidates.
Standards
IEC 62304:2006+A1:2015 — Software lifecycle ISO 14971:2019 — Risk management ISO 13485:2016 — Quality systems IEC 62366-1:2015+A1:2020 — Usability IEC 81001-5-1:2021 — Cybersecurity FDA Premarket Cyber (2025)
Regulatory Reviewer
Catches submission gaps, flags incomplete intended use, predicate issues, SOUP register holes
QA Auditor
ISO 13485 compliance — CAPA records, complaint trending, document control
Safety Engineer
Foreseeable misuse, hazard chain completeness, AI/ML automation trust
Cybersecurity Analyst
SBOM validation, threat modeling per AAMI TIR57, Section 524B
Clinical Reviewer — Neonatal SpO2
Domain specialist modeled on published literature (Bonafide et al., AAP consensus). Reviews SpO2 logic, alarm management, triage accuracy, nurse handoff quality.
SaMD-OS case study →
03
Clinical Evidence Design

Zero connected strength training device studies in the literature — evidence exists but no product bridges it.

Resistance training is first-line therapy for sarcopenia (ICFSR 2018, strong recommendation). Supervised adherence runs 72% — remove supervision and it drops to 43%. Functional outcomes respond (SMD 0.81–1.28) but muscle mass doesn't. Across 34 digital interventions for 60+, zero use a connected strength device.

Insight
Function, not mass, is the measurable outcome. Grip strength, gait speed, and TUG improve with progressive loading (40–80% 1RM). Muscle mass measures (SMM, ASMI, LLM) are all non-significant. The real constraint isn't efficacy — it's adherence after supervision ends, and no connected device addresses the supervision gap.
What I Built
Clinical Pathway Designer: Provider workflow + patient journey connected by a shared device data spine — load progression, session compliance, baseline strength. Illustrative FHIR CarePlan with functional-only goals. 7 evidence anchors (E1–E7) tracing every design decision to published literature. Every phase maps to ICFSR barriers: cost, transportation, lack of support.
Clinical Pathway case study →
Other Work
Draft Day
Real-time NFL draft companion. 50+ users, zero downtime.
QB Prospects
141 QBs scored, compared, and scouted. Search by name or archetype.
Nourish
Meal planning with macros, grocery lists, and Claude-powered recipe search.
flash_crash
iOS trading game. Shipped through App Store.