Synthetic encounters
Complete specialty encounter transcripts, SOAP notes, ICD-10/CPT codes, and prior-auth narratives. [ASPIRATIONAL]
Product · Synthetic clinical data
Augur —Training data without PHI risk.
Augur generates statistically faithful synthetic specialty encounters and medical images using NeMo generative LLMs and Clara/MONAI diffusion models trained on Scribara's de-identified corpus — covering the rare cases that real datasets under-represent.
Layer 4 — synthetic training fuel for the Scribara flywheel and external healthcare AI
NeMo generativeClara / MONAIDGXFDA-readyHIPAA safeNVIDIA Inception
NeMo encounter LLM7B param
Clara imaging diffusionDICOM
Clinical accuracy validatorNeMo
Privacy auditMIA test
The problem
Why Augur exists
Healthcare AI companies can't train on real PHI at scale; rare specialty presentations are chronically under-represented in real datasets. Existing synthetic tools generate structurally plausible but clinically shallow records that fail specialty accuracy checks.
- NeMo 7B+ parameter generative LLM trained on specialty encounter corpus [ASPIRATIONAL]
- Clara/MONAI diffusion models for derm lesion, retinal, and X-ray synthesis [ASPIRATIONAL]
- NeMo clinical accuracy classifier validates every sample before release
- Membership-inference attack audit confirms no real patient memorization
01 / Train
A generative model on specialty data.
NeMo LLMs fine-tuned on Scribara's de-identified encounter corpus learn the statistical patterns of real specialty medicine — diagnosis distributions, coding conventions, prior-auth language. [ASPIRATIONAL]
02 / Generate
Encounters and images on demand.
Parameterized generation: specialty, diagnosis category, complexity, demographic. Clara/MONAI produces synthetic DICOM images that pass clinical realism checks. [ASPIRATIONAL]
03 / Validate
Clinically correct, not just plausible.
A NeMo classifier trained on specialty coding logic rejects samples where the diagnosis, codes, and clinical narrative don't cohere — ensuring medical accuracy, not just structural validity.
04 / Deliver
Internally and externally.
Rare-case coverage flows to Forma. External customers — healthcare AI startups, research institutions, FDA testers — get bulk exports or API access to validated synthetic datasets.
Capabilities
What Augur produces
Synthetic imaging
DICOM-format derm lesions, retinal fundus, and ortho X-rays from Clara/MONAI diffusion models. [ASPIRATIONAL]
Clinical accuracy validation
NeMo classifier rejects implausible samples before they reach any customer or training pipeline.
Privacy audit
Membership-inference attack testing confirms no real patient data is memorized or reproducible.
Forma flywheel fuel
Rare-case synthetic coverage fills the long tail of presentations Forma's real-data stream under-represents.
External API
Healthcare AI startups and research institutions purchase validated synthetic datasets via bulk export or generation API. [ASPIRATIONAL]
GPU-essential
Why DGX-scale compute is required
Training a 7B+ parameter NeMo encounter LLM requires 8–64× H100/H200 GPUs; there is no CPU path. Clara/MONAI diffusion model training is similarly GPU-bound. Production inference — generating 10,000 synthetic encounters in a batch — requires TensorRT-optimized GPU inference.
- NeMo 7B+ LLM training: 8–64× H100/H200, DGX [ASPIRATIONAL]
- Clara/MONAI diffusion training: GPU-bound, no CPU path [ASPIRATIONAL]
- TensorRT generation API: batch synthesis 100–1000x faster than CPU
- RAPIDS validation analytics: corpus distribution checks at scale
DGX training clusterH100/H200
NeMo LLM7B+ params
Clara/MONAI diffusionimaging
TensorRT API<5 s/image
Impact
What it moves
0
Parameter encounter LLM [ASPIRATIONAL]
0
Per synthetic image on GPU
0
Specialty encounter models
0
Global synthetic health data market (2026)
The ecosystem
Works with the rest of Scribara
Augur feeds synthetic training data to every product that needs rare-case coverage.
Answers
Frequently asked
Augur includes provenance tracking and privacy audit trails that position its datasets for FDA AI/ML SaMD testing use cases. Specific regulatory acceptance [UNKNOWN] — confirm with your regulatory counsel.
Augur trains on de-identified Scribara encounter outputs; every generated sample passes a membership-inference attack test confirming no real patient data is memorized. PHI never enters the generative pipeline.
Yes — Augur will offer bulk dataset exports and a generation API to healthcare AI companies, research institutions, and regulatory testing organizations. Pricing [UNKNOWN] — contact us. [ASPIRATIONAL]
Train on data that doesn't expose patients.
See Augur generate a batch of specialty encounters and validate their clinical accuracy.