Owned ASR
Specialty acoustic + language models on Riva/NeMo, edge and cloud parity. [ASPIRATIONAL]
Platform
Intelligence Scribaraowns — end to end.
From the microphone in the exam room to the claim in the clearinghouse, Scribara runs its own models on NVIDIA compute — trained on a compounding clinical dataset and deployable inside your datacenter.
One platform across perception, reasoning, action, and learning
Epic FHIRathenahealthNVIDIA InceptionHL7SOC 2HIPAA
01 / Perceive
Speech and vision, owned.
Vocalis (Riva) hears the room; Visera (Clara/MONAI) sees the findings. Specialty-tuned, on-prem capable, confidence-scored.
02 / Reason
Specialty clinical reasoning.
A NeMo model drafts notes, codes, and justifications grounded in guidelines and payer rules through a knowledge graph and NeMo Retriever.
03 / Verify & act
Guardrailed autonomy.
An independent verifier and NeMo Guardrails enforce abstention; the agent acts only with clinician sign-off, then logs an immutable trail.
04 / Learn
A compounding flywheel.
Forma turns every edit into labeled data and continuous fine-tunes on DGX — the moat strengthens monthly.
Capabilities
What the platform provides
A complete, governed substrate for clinical AI — not a model endpoint.
Accelerated serving
TensorRT-LLM + Triton multi-model serving under a 300 ms in-exam budget.
On-prem inference
NIM microservices on NVIDIA AI Enterprise — PHI stays in your boundary. [ASPIRATIONAL]
Model foundry
Per-specialty and per-provider fine-tunes with eval-driven CI (Forma).
Clinical RAG
NeMo Retriever over a guideline + payer-rule knowledge graph.
Governed by default
Guardrails, abstention, audit, and per-tenant isolation built in.
GPU-essential
Why accelerated compute is the substrate
Training the ASR, reasoning, verifier, and imaging models is GPU-bound; real-time multi-model inference and on-prem deployment require it too. GPUs are not a nice-to-have here — they are the product.
- Training on DGX / H100 / H200 / GB200
- Real-time inference via TensorRT-LLM + Triton
- RAPIDS pipelines over 100M+ encounter records [ASPIRATIONAL]
- Edge inference on Jetson Orin + Holoscan [ASPIRATIONAL]
DGX training clusterfine-tune
Triton servingmulti-model
RAPIDS ETL100M+ rows
Jetson edgeon-prem
Data flow
From microphone to claim
A single governed pipeline carries the encounter through perception, reasoning, verification, and action — then back into training.
- 1
Capture
Audio and images enter; PHI is redacted at ingress.
- 2
Understand
ASR + imaging produce structured, confidence-scored findings.
- 3
Compose
The reasoning model drafts note, codes, auth, and referral.
- 4
Verify
The verifier checks; low-confidence spans abstain or escalate.
- 5
Write & learn
Clinician approves; EHR write-back; edits flow to Forma.
Developers
Built on open clinical standards
Epic FHIR and athenahealth integrations, idempotent write-back, and an immutable audit of every action.
# Sign and post a completed encounter to the EHR
POST /fhir/r4/DocumentReference
{
"status": "current",
"type": "specialty-soap-note",
"codes": ["I25.10", "93306"],
"verifier": "passed", "signed_by": "npi:1639…"
}
Performance
Engineered for the room
0
In-exam inference budget
0
Uptime target [ASPIRATIONAL]
0
Encounters in the flywheel [ASPIRATIONAL]
0
Owned model families
Trust
Enterprise-grade from day one
Healthcare buys on trust. Scribara is built to clear procurement, security review, and compliance.
HIPAA / BAA
PHI handled under a Business Associate Agreement.
SOC 2 Type II
[ASPIRATIONAL] — roadmap to Type II within 12 months.
On-prem NIM
Run owned models inside your datacenter via NVIDIA AI Enterprise. [ASPIRATIONAL]
Immutable audit
Every agent action logged, reversible, tied to the clinician's signature.
Data residency
US, EU, UK regions; per-tenant keys (BYOK option). [ASPIRATIONAL]
ISO 42001
AI management-system conformity on the roadmap. [ASPIRATIONAL]
Answers
Frequently asked
Yes — the models package as NIM microservices on NVIDIA AI Enterprise, so they run inside your datacenter with PHI resident. [ASPIRATIONAL]
Forma produces per-specialty and per-provider fine-tunes from your de-identified corrections, with eval-driven CI before anything ships.
ASR, reasoning, verifier, and imaging are owned and NVIDIA-served; frontier APIs are a bootstrap/fallback only.
See the platform run an encounter.
A technical walkthrough of perception, reasoning, verification, and on-prem deployment.