The Official Record Is Late

At Epic, an alert could be delivered and still fail to help a clinician. At Doximity, a verified physician account did not solve the problem of what number a patient saw. At Andwise, physicians signing up did not settle who would pay.

These were different problems. Each required looking beyond the record the system already kept to the decision someone still had to make.

RNA research

My work began in Roger Wartell’s lab at Georgia Tech, studying how Hfq affects small regulatory RNAs and their pairing with target mRNAs. I was first author on an ACS Symposium Series chapter on that work, and later researched cancer-derived exosomes at Memorial Sloan Kettering.

That work taught me to test a proposed explanation against observed behavior. RNA folding is a molecular process; it does not establish a mechanism for institutional failure.

Alerts

At Epic, I worked on EHR configurations. Decision-support alerts intended for patient safety could overload clinicians until they dismissed them. A record saying “alert delivered” did not tell us whether the alert could be acted on at that moment.

The questions were about the workflow: how often the alert fired, which alerts needed immediate attention, and what a clinician could do with them.

Dialer

At Doximity, I originated Dialer as founding product lead. A physician could call a patient from a personal phone while the patient saw the office number.

Twilio required proof of access to the displayed number. We used a vendor that did not, and I built a working prototype over a weekend. Doximity had verified the physician behind the call; the displayed number itself was unverified. It belonged to the practice or hospital, so the choice also involved their identity.

Physicians began using the product before hospital authorization, which came in stages. The case study records that decision and its limits. Doximity reports more than 300,000 calls on an average workday across 250+ hospitals and health systems in February 2026. Those figures establish scale; they do not isolate which design decision caused adoption.

At CancerCompass, I rebuilt dense oncology information around plain-language pathways and a route to nursing triage. At Transcarent, I directed four specialty-care programs and replaced manual case coordination with routing tied to clinical roles.

The practical question was whether a patient reached the next step, with someone responsible for it. An app open or delivered message could not answer that by itself.

Andwise

I co-founded Andwise and built software to flag clauses in physician contracts for review. We raised $240,000 and reached 1,200 physician users.

I left the payer question until year two. Sustained engagement and funding fell short, and I wound the company down in 2024. The case study describes the options we considered and the email announcing the wind-down.

Transition forecasting

NextConsensus addresses a timing question: how likely is a named body to change a guideline, label, or coverage decision by a team’s deadline?

Each forecast freezes its evidence cutoff, probability, baselines, and resolution rules before the outcome. It can then be scored against the record. No registered forecast has resolved yet, so there is no accuracy record to report.

This is where the title applies directly: a published policy can describe yesterday’s position while new evidence is under review. Predicting the next decision requires testing; the earlier episodes do not prove that the forecast works.

Corrections

7 October 2026: this essay said “Over two decades.” The arc it describes runs from 2009 to 2026, seventeen years.
7 October 2026: this essay said I “spent a summer on thoracic surgery outcomes at Memorial Sloan Kettering.” The internship was research on cancer-derived exosomes and the premalignant changes linking lung cancer and COPD.
7 October 2026: this essay’s diagram and closing said that in all three examples “the true state moves first and the record follows.” Only the formulary example is a delay; in the RNA and alert examples what the record says and what is happening come apart without one following the other.
7 October 2026: this local edition was abridged and rewritten. It removes the diagram and the claim that RNA folding, alert fatigue, and policy changes are the same structural problem, and separates Dialer’s reported scale from a causal explanation of adoption. The original edition remains the version of record.