Most systems I have worked inside have an official record that lags reality. In healthcare, the lag can kill people.
I came to it by building inside those systems. Over two decades, layer by layer, the same structural failure kept surfacing in different forms. This essay makes the arc visible.
RNA folding
At nineteen, I was in Roger Wartell’s lab at Georgia Tech, modeling the thermodynamic barriers that determine whether a gene-regulation reaction proceeds or stalls. A strand of RNA could carry the right nucleotides and still fail — if the folding produced a barrier that blocked interaction. The sequence was correct. The shape was wrong. Structure determined what the information could actually do.
I co-authored a book chapter in the ACS Symposium Series on sRNA-mRNA interactions and Hfq, spent a summer on thoracic surgery outcomes at Memorial Sloan Kettering, and absorbed something I could not yet articulate. Most consequential failures live in the distance between a system’s official design and its actual behavior. Who can act on information, under what constraints, with what review path, on what timeline — these determine whether the information matters at all.
Alerts
After Georgia Tech, I implemented EHR configurations at Epic. Decision-support alerts designed for patient safety overloaded clinicians until they reflexively dismissed them. The alerts were correct. Too many, wrong cadence, no triage between binding and advisory. Clinicians ignored the system because the structure had made compliance impossible. The official record said “alert delivered.”
Dialer
At Doximity, I was founding product lead for Dialer — a product that let doctors call patients from personal devices while the patient’s phone showed the office number. The structural problem: physicians need to reach patients but can’t share their personal number — once it’s out, it can’t be un-shared, and the patient calls back at 2am. So they call from the hospital desk phone, which means they’re at the hospital, which means the call happens during hospital hours or not at all. Most choose not at all.
The shortcut
Twilio required verifying access to the office number before letting us display it. We found a vendor that didn’t, and over a weekend I built a working prototype placing outbound calls under an unverified caller ID.
I knew that approach would never pass hospital IT security review — but I also knew hospital IT would never approve the product on paper anyway. It lets physicians use their office number from a personal cell phone, geographically liberating the number from the hospital phone system. No security team green-lights that before it’s in use.
The only path through was around: ship to physicians, build critical mass, let approval follow adoption. A doctor running between rooms. A patient burned by spam. A missed connection that could be a missed diagnosis. The signal had to survive the worst day. Dialer now carries more than 300,000 calls on an average workday, across 250+ hospitals and health systems, because the patient sees the office, not “spam risk.”
Corrected 20 August 2026: this paragraph previously read “300,000+ daily visits across 200+ health systems — pickup rates 3x higher than competitors.” The call volume and the health-system count are from Doximity’s February 2026 investor release, which says calls rather than visits and 250+ rather than 200+. I have not been able to source the 3x pickup-rate comparison, so it is removed rather than repeated.
The middle
Across oncology navigation for 30 million annual visitors and care coordination programs spanning surgery, urgent care, behavioral health, and oncology — the same thing. The official record lagged. Fragmented guidance made every case an exception. Metrics built around app opens instead of completed care pathways. The structure and the behavior were at war.
Andwise
Co-founding Andwise — a physician financial wellness company — brought the pattern to a new surface. Physicians navigating contracts, debt, home buying, and the transition from training faced the same structural problem I had seen in clinical systems: fragmented guidance, trust-sensitive information, no review path when something went wrong.
We raised $240,000, grew to 1,200 physician users, and built a medical advisory board of over fifty physicians. The constraint map became the operating manual: which decisions were binding, which could be adjusted, which required escalation, what the response window was for each. The re-review trigger fired whether anyone remembered to check.
Transition forecasting
Now the object is medical knowledge itself.
NextConsensus forecasts institutional action from medical evidence. Coverage determinations, formulary placements, guideline updates, and regulatory decisions depend on evidence that shifts continuously — new trial data, safety signals, label updates, and expert convergence. Most organizations find out when an authority decides, not when the evidence turned. By the time a committee convenes, the recognition window has already closed.
NextConsensus registers probability forecasts for defined authority transitions, freezes the evidence state at registration, and scores every outcome against public resolution rules. The pipeline is reproducible. The evidence cutoff is immutable. The forecast can be scored because the evidence was frozen before the outcome.
The same shape
A strand of RNA carries the right sequence but folds into a shape that blocks function. An EHR alert is correct but fires too frequently for anyone to process. A formulary position was accurate last quarter but the evidence has moved. The official record is late. Same structural problem, different scale.