DxFlow harnesses LLMs for adaptive history-taking and clinical information gathering, then applies physician-curated, deterministic pathways to generate a focused differential and flag potentially dangerous diagnoses. Each recommendation can be traced to the findings and logic behind it. The clinician makes the final decision.
Differential diagnosis
Medical decision-making is hard — it takes sound reasoning and consistency, under pressure, on every patient.
The diagnosis usually lives in the history — but taking a thorough, pertinent one on every patient, under time pressure, is hard.
The time-critical, can’t-miss conditions have to be considered in every scenario — and they’re the easiest to overlook.
Committing to a plan takes reasoning you can trust — backed by medical evidence and years of experience, not guesswork.
A general LLM is a brilliant generalist — but it can confidently give a plausible-but-wrong answer and change its mind each time you ask. Medicine needs a purpose-built (vertical) system, not a generalist.
DxFlow works the way a careful clinician does — one focused question at a time, always thinking about what could be dangerous.
Begin with the chief complaint. The intake follows traditional medical teaching for that scenario, and every next question is high-yield — each one takes you a step closer to the diagnosis.
A ranked list of diagnoses with odds updates as you answer — and the can't-miss causes are always checked first.
DxFlow recommends the next best test or action — the right test, not every test — and shows the reasoning, so you decide with a clear picture.
Large language models (LLMs) like ChatGPT are remarkable generalists. But a general LLM predicts the next word from patterns across the whole internet — so in a high-stakes moment it can give a plausible-but-wrong answer, phrased as confidently as a right one, and a different answer each time you ask. DxFlow is built the opposite way: a vertical system where the reasoning is computed, the content is physician-authored, and every number traces back to a source.
| General LLM | DxFlow | |
|---|---|---|
| Reasoning | Generative, token-by-token | Deterministic, calibrated engine — reproducible |
| Content | Model training data | Physician-authored banks & guidelines |
| Safety | No can't-miss guarantee | Dangerous-first; designed not to drop a can't-miss diagnosis |
| Evaluation | General benchmarks | Derivation-graded on real cases & physician-vetted |
| Output | Freeform prose | Differential + odds + work-up + citations + note |
| Auditability | Opaque | Every % traces to a physician-authored criterion; answers cite the source |
Same clinical question, same answer, every time — with the dangerous diagnoses checked first and the reasoning shown, not asserted. The difference between a generalist that talks about medicine and a vertical system built for it.
Built to be trusted, not just believed — every answer is grounded, checked, and traceable.
Every finding, diagnosis, and care pathway is written and vetted by a practicing physician — not scraped from the open web.
The reasoning is graded against real, de-identified cases — did it surface the dangerous diagnosis, ask the basics first, capture what matters — not just the final answer.
The same case yields the same reasoning every time — reproducible, not a fresh guess on each visit.
Every probability and recommendation traces back to a source you can check. Decision support — you make the call.
Clinical decision support — designed to assist, not replace, clinical judgment. Not a diagnosis, and not FDA-cleared.
Physicians, nurse practitioners, and physician assistants across primary care, urgent care, emergency, and hospital medicine — an experienced reasoning partner at the point of care, especially when there's no senior colleague to turn to.
Practice structured reasoning and build the "dangerous-first" discipline, with the logic visible at every step.
No patient names are ever stored — cases are de-identified, so it's safe to use in everyday practice and teaching.
Short bio — a few lines to add
I've been in clinical practice for over 25 years, and I've made my share of mistakes. To this day I still dread misdiagnosing a case, or delaying a diagnosis and the treatment that depends on it. Medicine has progressed tremendously — MRI and CT scans, endoscopy, and a whole host of lab tests now help us reach a diagnosis. But one thing hasn't changed: the art of history taking — listening to the patient for the clues to the diagnosis.
I think of a patient who arrived with sudden abdominal pain, nausea, and diarrhea. The first explanation — gastroenteritis — seemed to fit. But something didn't: the pain was far more severe than the exam suggested, and a more deliberate history uncovered atrial fibrillation and a recent lapse in anticoagulation. That changed the differential. Instead of sending the patient home with fluids and reassurance, we pursued urgent imaging and caught intestinal ischemia before the window to act had closed.
Unfortunately, I see many of us moving away from that art and rushing to order all sorts of tests. It drives unnecessary cost and, too often, inaccurate diagnoses. Many colleagues are also turning to ChatGPT for answers. It's helpful, but inconsistent — and at times it gives grossly inaccurate, even invented, answers (technically, "hallucinations") that can be very costly. In medicine we can't allow that, even if it's rare.
So I built DxFlow: a tool that thinks the way I was trained to — dependable and deterministic. Its purpose isn't to hand you the final answer right away, but to build a reasonable differential and, above all, to make sure life-threatening conditions are never missed. It helps me order the right tests — stopping the unnecessary ones, and prompting the ones I might otherwise have overlooked.
I built it for everyone who carries that responsibility — physicians, and especially the nurse practitioners and physician assistants so often on the front line, seeing undifferentiated patients across every system, frequently without a senior colleague to turn to. DxFlow is meant to be that experienced colleague in the room, so you can act with confidence and know that you are doing what anyone else will do in that situation.
Every diagnosis, finding, and judgment in it is curated and corrected by a physician — not scraped from the internet. It's the experienced reasoning partner I wish I'd had as a resident, and the one I still want on a busy day. And it doubles as a learning tool for trainees.
Request access to DxFlow. Accounts are reviewed and approved so the community stays clinician-only.
Request access Clinicians verify with an NPI · Students welcome · You'll be notified once approved