Determinism
The same biometry and Doppler produce a byte-identical plan — yesterday, today, and five years from now. Across 50 configurations × 1,000 seeds, all 50,000 runs were bit-identical.
The mission
Better than AI — because it has to be.
ObGynAssist™ was not built to be faster or flashier than a chatbot. It was built to be more reliable, more traceable, and more correct — the same input always yielding the same answer, every threshold wired to a named guideline, and every decision open to audit, rule by rule. In obstetrics, that isn't a luxury. It's the baseline.
Why we built it
An engine, not a chatbot.
A general language model can sound authoritative and still be wrong in ways no clinician can see. For fetal growth, that's disqualifying.
The appeal of a conversational assistant is obvious — ask a question, get fluent prose. But underneath, a language model samples from a distribution. Ask the same question twice and the answer can change. It can invent a cut-off that no guideline ever published, attach a citation that doesn't exist, and offer no way to replay how it reached either. When the stakes are a missed growth restriction or a mistimed delivery, fluency is not the same as correctness.
So we took the opposite path. Instead of fine-tuning a model, we built a deterministic rule engine: input goes in, and a sourced classification comes out through a fixed pipeline that is the same every time it runs. Each threshold names the guideline behind it. Every output carries its engineVersion, a full-content rulesetSHA, and the exact firedRules[] that produced it — so any result can be replayed, bit for bit, years later.
It is not a black box that happens to be confident. It is a system that can show its work.
We didn't make a model
sound like a clinician.
We made an engine you can
audit like a textbook.
Where a chatbot offers a paragraph, the engine offers a traceable record: a classification label, the guideline citation behind it, and the precise rules that fired — reproducible on demand.
Deterministic
Sourced
Auditable
On-device
Design principles
Five commitments, built in — not bolted on.
Every part of ObGynAssist follows from a small set of non-negotiable principles. They shape the architecture, not just the copy.
Sources
Each decision names the reference behind it — RCOG GTG, NICE NG, ISUOG, ACOG, INTERGROWTH-21st, Delphi 2016 and more — backed by 58 guideline reference cards.
Traceability
When a plan is questioned, the audit trail shows precisely which step fired which rule. firedRules[] plus a full-content rulesetSHA make every output replayable.
On-device by architecture
The engine runs entirely on the iPhone — no network call, no cloud round-trip. A full sourced plan computes in real time, with patient data staying on the device.
Non-directive by design
Outputs are classification labels and citations that support clinician judgment — a companion, never a directive. Acuity tags are tags, not orders.
PROVENANCE ON EVERY OUTPUT
Every result is stamped with its engine version, a full-content ruleset hash, and the rules that fired — the structural guarantee that makes the four principles above verifiable rather than aspirational.
How the engine works →
Positioning
A companion, not a device.
ObGynAssist is a clinical companion and a Medical Reference tool — not a regulated medical device. Its outputs are classification labels and guideline citations that support clinician judgment; they do not replace it.
That framing is a deliberate design choice, and it runs all the way into the words on screen. The product follows a strict companion-language discipline: it describes findings as consistent with, associated with, or criteria met for a classification — and it surfaces acuity tags such as ADMIT · CONSIDER as labels for the clinician to weigh, never as instructions to follow. Where a clinical question remains open, the engine says so: clinician evaluation required.
The result is decision support that stays in its lane — honest about what it is, and honest about what it is not.
THE DISCIPLINE, IN PRACTICE
- →Findings framed as consistent with a classification.
- →Acuity tags are classification labels, never orders.
- →Open questions returned as clinician evaluation required.
- →Dose-band classification as scope — never a directive to act.
Medical Reference
Non-directive
Support, not replace
ObGynAssist™ supports — never replaces — the judgment of a qualified healthcare professional. No doctor–patient relationship is created by its use.
Research · in progress
A two-paper program.
Credible decision support earns trust the slow way — through evidence. Two studies are underway: one describing the system, one designed to test it against specialist assessment.
The methods & architecture paper
A description of the deterministic, versioned, rule-traceable design — how input becomes a sourced classification through a fixed pipeline, how provenance is stamped on every output, and how the system is verified. Intended for a medical-informatics journal.
Methods
Architecture
Reproducibility
The prospective clinical-validation study
A planned study comparing the engine's classifications against specialist assessment on real obstetric cases. This is the study that would move the evidence base beyond synthetic concordance — it is in design, not yet run.
Prospective
Specialist comparison
Planned
What is — and isn't — measured today.
To date, the engine has been graded against an independent guideline oracle on synthetic cases. That is agreement with standard-of-care references — a measure of internal consistency and guideline fidelity — and it is explicitly not a measure of neonatal outcomes. The prospective study is being designed precisely to close that gap.
0%
Oracle concordance · 2,994 graded synthetic cases
0
CI-locked oracle v1.0.0 · 1,000 synthetic cases
50,000 / 50,000
Bit-identical runs · determinism, proven
No publication, peer-review acceptance, author roster, or clinical result is claimed here. Concordance figures are measured on synthetic cases against guideline references, not on neonatal outcomes. Read the full validation →
One umbrella
ObGynAssist™ — and its modules.
ObGynAssist is the umbrella. Beneath it sit three module families, each a plain-text name for a self-contained, guideline-grounded body of work.
FetalGrowth
The flagship module: FGR / SGA / AGA / LGA staging, Doppler analysis, and the twin spectrum across DCDA, MCDA and MCMA.
RCOG · ISUOG · NICE · Delphi 2016PregAssist
Eight pregnancy-assist modules — dating, ectopic, aspirin and VTE prophylaxis, thyroid, hyperemesis, ASA and CMV — each self-contained.
RCOG GTG · NICE NG · ISUOGGynAssist
The gynecology family, beginning with recurrent pregnancy loss — an evidence-based workup with guideline-gated logic.
ESHRE 2022 · RCOG GTG 17Ten modules in all, one engine, one provenance discipline — the same determinism, sourcing and traceability whether the question is a growth-restricted twin or a recurrent-loss workup.
10 modules
8 pathways
243 rules
Get in touch
Build the evidence with us.
ObGynAssist is coming to the App Store as a Medical Reference companion. For early access, clinical partnership, or a conversation about the prospective study, get in touch.
Coming to the App Store · Medical Reference