McCay Barnes — Delightful design with clinical integration, hardened through design controls

Selected Work · 2026 / McCay Barnes

Delightful design with clinical integration, hardened through design controls.

LIVE TRACEEX-PREEMIE · MILD BPD · BASELINE 91% · ILLUSTRATIVE DATA NOW · 21:00

21:0000:0003:0006:0009:00

/ Selected Work

Claim 01

Warm handoffs at data scale.

Some patients need a clinician's attention but never get it. Not because the model is wrong, but because no human can review every overnight SpO₂ trace or call every post-cataract patient daily. AI absorbs the data volume that was hiding those patients, then delivers a warm handoff a clinician can act on. Two domains, same posture: NICU monitoring that surfaces the overnight trace a neonatologist needs to see; post-cataract symptom check-ins that route escalations into same-day call-backs with the transcript pre-loaded. Both examples are AI evals. The pipeline and scoring are real, based on synthetic patient data.

SpO2 · LLM-as-Judge Eval 90 / 100 / 100 final

V1 · pre-fixes 80% 90% 100%

V2 · regression 70% 30% ↘ 100%

V2.1 · after fix 60–70% 80% 100%

V3 90% 100% 100%

Lead Evidence

SpO2 Eval Pipeline

Overnight neonatal SpO₂ traces turned into warm clinical handoffs, not cold alerts. Three LLM judges scored clinical accuracy, handoff quality, and artifact handling across four iterations (10 traces, 30 evals each). V2 regressed handoff quality from 90% to 30%. The eval caught it. V3 (2026-04-29): 90 / 100 / 100, and 0 urgent false negatives across 400 traces.

Open Project ↗

Cataract · LLM-as-Judge Eval 92% sensitivity

Mrs. Chen, day 7 — redness + new floater + haloes Escalate

Patient K, day 2 — severe disproportionate pain Escalate

Patient G, day 3 — single tiny floater Pass

Patient F, week 1 — routine recovery Pass

Patient B, week 4 — asymptomatic PCO Missed

Second Evidence

Cataract Triage Eval

Between-visit patient symptoms turned into same-day clinical escalations. A conversational agent runs a 5-symptom interview (Wanten 2025), classifies each response, and on escalate hands the clinician a same-day call-back with the patient's transcript pre-loaded, not a blank intake form. 12 of 13 should-escalate cases caught. The miss, week-4 asymptomatic PCO, sits in the 92%: symptom interviews can't catch structural findings.

Open Project ↗

Claim 02

Pathways no one's built yet.

Most new growth in regulated healthcare lives in pathways that haven't been built yet. The blocker isn't ideas. It's the cost of working one up to a fundable proposal: provider workflow design, patient journey mapping, evidence-to-design synthesis, illustrative care planning that survives clinical review. Months of work per concept. The skill that produced the sarcopenia pathway on this site does all four in an afternoon. What you get isn't a launch-ready product. It's a fundable proposal with real substance, end-to-end. The unlock is bringing 10 of these to a budget conversation instead of one. The decision shifts from "should we bet on this" to "which of the ten is worth the deeper investment."

Clinical Pathway · Evidence Base E1 → E7

E1 · guideline E2 · efficacy E3 · adherence E4 · digital E5 · guideline E6 · barriers E7 · long-term

Method · Worked Example

Clinical Pathway Designer

Sarcopenia is the worked example. Two workflows (provider + patient) sharing one device data spine, one illustrative FHIR CarePlan, anchored in seven literature citations. The pattern transfers; any condition with a connected-device thesis and curated evidence ships in the same afternoon.

See the worked example ↗

/ How I Work

Claim 01

Design controls in hours. Tied to code.

Design control drafts traditionally chase the code: weeks of asking dev for context, scheduling reviewer sessions, getting shallow drafts back. SaMD-OS inverts that. It generates design controls, hazard analyses, and SOUP registers directly from the codebase, anchored to pinned standards, then runs five specialist reviewers (regulatory, QA, safety, cybersecurity, clinical) in parallel. The most conservative verdict wins. Hours of detailed drafts, instead of weeks of shallow ones. The drafts feed your eQMS; the eQMS still owns the approved record.

SaMD-OS · Architecture 4 layers · 20 skills

Builder Skills

6 generators · PRD, design controls, risk, FHIR, change impact

Reverse Engineering

3 extractors · code → SOUP, inputs, hazards

Specialist Reviewers

5 red-team · regulatory, QA, safety, cyber, clinical

PM Skills

6 day-to-day · PRD, metrics, decisions, status, competitive, roadmap

Architecture Pattern

SaMD-OS

A working pattern for keeping IEC 62304, ISO 14971, and SOUP artifacts in sync with shipped code. Five specialist reviewers run in parallel against pinned standards; the most conservative verdict wins. The toolkit is the proof; the architecture is the point.

Open Project ↗

Claim 02

v5 in v1 time.

Working code is the easy part. The iteration happens before the first line of code: multiple Claude sessions stress-test the framing, the spec, and the plan, and specialist persona reviewers (clinical, regulatory, safety, cybersecurity) red-team the plan against pinned standards. By the time code lands, v5 of the idea is on the page. One live pass catches integration issues. What reaches your calendar is fully formed, not half-baked.

Pipeline · Pre-Code Iteration 4 stages

Ideate

Multi-session PM peer review · framing, assumptions, prior art

Spec

Multi-session PM peer review · scope, risk, success criteria

Plan

Specialist red-team · clinical, regulatory, safety, cyber

Live check

Integration review · visual + system-level catches

Method · orchestrate-build pipeline

Iteration before code

Most of the value lives in the cycles before code exists. Multiple sessions sharpen the framing, the spec, and the plan; specialist personas red-team the plan against pinned standards. Code is the last step, not the first. The cycles are the proof; the iteration density is the point.

See it run in SaMD-OS ↗

Claim 03

From client question to board case.

Data-product strategy sprawls across decks, memory, and one-off analyses, and every PM runs it differently. PM-OS makes it one versioned toolkit: a client use case runs through a data-feasibility verdict and an evidence-bar review, then a business-case engine sizes the opportunity and a right-to-win skeptic pressure-tests every number before the board sees it. The PM still owns the call; the data team still builds. Same playbook, every PM.

PM-OS · RWD Strategy 5 layers · 11 capabilities

Listen & Frame

3 skills · ideate, concept-gen, pm-advisor

Scan

2 skills · competitive intel, positioning teardown

Validate

2 skills · data feasibility, evidence review

Make the Case

the board engine · sizing, skeptic, deck

RWD Strategy Toolkit

PM-OS

A working pattern for turning a client question into a board-ready case — a data-feasibility verdict, an evidence-bar review, and a business case pressure-tested by a right-to-win skeptic before leadership sees it. Run the same way by every PM. The toolkit is the proof; the playbook is the point.

Open Project ↗

Playground

01 Draft Day Real-time NFL draft companion. Fifty-plus users, zero downtime, which is the kind of stat that only matters if you have ever lost a draft to a crash. Visit ↗ 02 QB Prospects 141 quarterback prospects, scored and compared and searchable by name or archetype, which is the kind of thing the draft media keeps not building. Visit ↗ 03 Nourish Meal planning with macros, grocery lists, and recipe search, built end-to-end the way the existing tools never quite manage. Visit ↗ 04 flash_crash An unbeatable iOS game like Flappy Bird except with stock trading graphs. Emphasis on motion algorithms, playability per skill type, and design aesthetics. Visit ↗