▸ THE BUILD · CODEBLUE PEDS · IN PRODUCTION

Your protocols. Their pocket. Seconds.

At 2 AM in a rural ED, the clinician caring for a deteriorating child needs their institution's own validated protocol in hand in seconds — not a login, not an app, not a generic chatbot. CodeBlue Peds turns that validated library into an SMS line any clinician can text.

~146

protocols live

~1%

confident-wrong rate

review cycles

doses the AI states

Measurements from the production system. Retrieval accuracy measured on an independent held-out test — 2 of 195 held-out cases, 95% CI 0.3–3.6%; method described under Accuracy below. Dose suppression enforced in code and verified by the test suite.

CB CodeBlue Peds PICU line · SMS

8-week-old, mottled, cap refill 4 sec — what's the pathway?

PHI STRIPPED · MATCHED · REVIEWED

Match: Neonatal Septic Shock Pathway (your institution, v3.2). Follow the pathway's assessment and escalation steps — full protocol: cbp.link/nss-32. For dosing, see the protocol's dosing table.

Not a substitute for clinical judgment.

Simulated exchange · grounded in the institution's own protocol

THE GAP

Pediatric expertise is concentrated in academic centers. Children get sick everywhere.

4 in 5

U.S. emergency departments are not fully prepared to care for children.Weyant et al., Health Affairs, 2024

9 in 10

Completely rural U.S. counties have no general pediatrician at all.Ramesh & Yu, JAMA Network Open, 2023

24M

U.S. children lack timely access to high-readiness pediatric emergency care.Joseph et al., Journal of Pediatrics, 2025

The problem was never the content — it's access. Institutions invest heavily in writing and validating clinical protocols, then store them where no one can find them at the bedside. What CodeBlue Peds closes is not "more medical knowledge" — it is retrieval of the institution's own validated knowledge at the moment of need.

The pipeline

text → strip PHI → match → ground → review → guard → deliver

the message pipeline · grounded, reviewed, dose-suppressed

Text the question

Any phone, plain SMS. "8-week-old, mottled, cap refill 4s — what's the pathway?"

PHI stripped in code

De-identification before storage or AI. Numbers exist only as salted hashes.

Matched to their protocols

Semantic search over that organization's validated corpus — tenant-isolated.

Grounded answer

Generation constrained to the matched protocol's content — the model is not the primary source of truth.

Independent review

A separate model reviews every answer before delivery. The ungrounded path fails closed.

Guardrails in code

Dose statements stripped, emergency language detected deterministically — in code, not prompts.

Delivered in seconds

With a link to the full protocol document.

Fail safe, never silent

If confidence is insufficient, the reply is the protocol directory — never a guess, never silence.

THE DEFINING ENGINEERING DECISION

The most important line of code in the system is the one that refuses to state a dose.

Pediatric dosing is weight-dependent and unforgiving. Rather than ship a dose-checker that could look safer than it is, the system suppresses dose statements entirely — enforced in code, and the system refuses to start in production if it is disabled — and defers to the institution's protocol. The willingness to remove a capability that couldn't be made trustworthy is the point. The worst case is a wrong link, not a wrong dose.

Guardrails in code, not prompts

Prompt instructions are suggestions to a model; code is a guarantee. Dose suppression and emergency detection are deterministic code paths.

No model reviews its own work

Grounded answers pass an independent second-model review before delivery. The review gate fails closed.

Tenancy as a safety property

One institution's clinician can never receive another's protocol. The isolation boundary is verified against real PostgreSQL, not assumed.

Only validated content is served

Uploads land in a quarantined pending state. The publish step is the only code path to a servable protocol.

ACCURACY PHILOSOPHY

Top-1 accuracy flatters. The number that matters is how often the system is confidently wrong — and we publish it, then hand institutions the test to measure it on their own protocols.

On an independent held-out test, the confident-wrong rate sits at roughly 1% — 2 of 195 held-out cases, 95% CI 0.3–3.6% — with the remaining uncertainty resolving to safe abstention (the directory), not a wrong answer. Evaluation questions are written from the clinical presentation, never the protocol's name, and held out from anything that built the search index. Confusable pairs — DKA vs. HHS, SVT vs. VT, epiglottitis vs. croup — are scored explicitly.

SELECTED WORK · PUBLICATIONS & PREPRINTS

The deployment work, written up in the open.

Measurements from a live deployment — the access gap it addresses, and the cross-institutional data behind it. Preprints and datasets are public; manuscripts under review are listed without venue.

SUBMITTED

Resource-Aware Pediatric AI: Decision Support by Subtraction for the 5-Bed Rural Emergency Department.

Hsu B · 2026 · under review

SUBMITTED

The Pediatric Decision-Support Access Gap: How Often Do Pathways Lack a Public Analogue?

Hsu B · 2026 · under review

PREPRINT

The Pediatric Decision-Support Access Gap: A Measured Absence of Peer-Institution Pathway Analogues Across a Deployed Pediatric Library.

Hsu B · Zenodo, 2026 · doi:10.5281/zenodo.21212699 ↗

DATASET

The Access-Gap Reproducibility Bundle — Cross-Institutional Comparison Outcomes for 150 Deployed Pediatric Pathways (CodeBluePediatrics).

Hsu B · Zenodo, 2026 · doi:10.5281/zenodo.21198499 ↗

For health systems

Want your protocols in every clinician's pocket?

We're speaking with health systems about putting their own validated pediatric protocols on an SMS line clinicians already know how to use.

benson.hsu@celeritashealth.com ← Back to Celeritas