Accuracy01

100%

59 of 59 clinical scenarios passed field extraction against a fixed answer key. No partial credits, no hand-curation.

F1 · 9 SOAP fields02

0.978

Field-level F1 ranges from 0.92 to 1.000 across the nine fields. Per-field figures are in the schema table below.

Scenarios03

Each scenario judged twice, by models from two different vendors, with no divergence between them. A small set judged transparently, not a large one judged once.

VOL 2 BUILD v2.2 APRIL 2026 ● PRODUCTION BENCHMARK

Clinical documentation,
validated three ways.

Air-gapped speech-to-structured pipeline. Runs on hardware you own. Two independent LLM judges from different vendors evaluated every case. We publish the numbers because medical documentation is the one place you cannot trust a pitch deck.

Request a pilot See the methodology

Benchmark run on 59 validated clinical scenarios. Multi-language rollout on the 2026 roadmap — each new language releases on the same two-judge methodology before general availability.

What we ship that cloud scribes don't

Sovereign by architecture. Honest by benchmark. Shipping today.

Cloud medical scribes solved a demo. They did not solve the compliance, the residency, or the cost of putting AI into a regulated clinical workflow. We built for that from day one.

Air-gap capable

Entire pipeline — STT, PII masking, LLM extraction — runs inside your hospital's data centre. Clean-slate regression validated. Audio processed in RAM only; no raw audio persisted to disk.

air-gap validated v0.4.5

Full-stack control

No vendor API in the hot path. No closed-model inference. Every inference runs on hardware you control, behind weights you can fingerprint per release. When your InfoSec team asks what executed behind a field, we answer with a hash, not a marketing page.

auditable · reproducible · hash-pinned

Two-judge benchmark

Every scenario evaluated by two independent LLM judges from different vendors. Structural, keyword, and semantic validators per case. One number is a marketing number; two independently-judged numbers with an audit trail are a benchmark.

two judges · no divergence

Guideline compliance

The gap you can still close — before the note is signed.

Extraction accuracy is table stakes. What clinicians ask for is to see, during the visit, which guideline-required steps are not yet reflected in the record. Every cloud scribe writes the note. None of them checks it against the guideline while it can still be corrected.

Checked against the guideline

For the coded diagnosis, Scribe returns a verdict per criterion: documented, missing from the record, or not evidenced in what was said. The clinician reads a short list of gaps, not a wall of generated text.

verdict per criterion

Your protocols, not only national ones

The same ingest engine takes any body of rules: national clinical guidelines, your own internal protocols and quality standards, or an insurer's requirements. That is what makes the product portable across markets instead of tied to one country's content.

tenant-scoped content

Prescriptions screened

Prescriptions are matched against the drug reference and screened for interactions between the drugs prescribed in the same visit. Contraindication screening linked to the coded diagnosis is in development and is not part of what ships today.

interactions · diagnosis link in development

Every verdict links back to the exact phrase that produced it. When a reviewer asks why a field says what it says, the answer is a quote from the consultation, not a model explanation.

Pipeline

Dictation in. Structured chart out. Nothing leaves the building.

Five stages. Every stage runs on hardware you control. Every stage is auditable. Every edit a clinician makes is versioned with a before-and-after diff.

Figure 1 Pipeline stages with wall-clock timestamps.

t+0 msstage 01

Dictation

Clinician speaks naturally — during or after the visit. Native desktop client, cross-platform, offline-capable.

desktop · offline

t+127 msstage 02

Speech-to-text

Domain-adapted medical STT with INT8 / FP16 quantisation. Sub-second partial transcripts; under 15 s for a full visit.

tuned on medical corpora

t+143 msstage 03

PII mask

Edge-based masking before transmission. Audio processed in RAM only. Structured output encrypted at rest per tenant policy.

no audio persisted

t+847 msstage 04

LLM extraction

Medical-domain LLM tuned for clinical field extraction. Runs on your GPUs. No vendor API in the pipeline.

in-house inference stack

t+12.4 sstage 05

Structured chart

Nine SOAP fields populated. Validation layer cross-checks extracted drugs against a reference of 2,830 substances, with 33,493 contraindications and 72,848 interactions.

clinician reviews & signs

Output schema

The nine fields Scribe fills.

Production schema defined in our transcribe routes. Field-level F1 scores reported from the April 2026 benchmark (v2.2 build).

Figure 2 Output schema with field-level F1 scores.

9 SOAP chart fields Scribe populates, with April 2026 benchmark field-level F1 scores
Field	What Scribe extracts	Field F1
Chief complaint	Presenting symptom(s) in the patient's own words.	0.98+
History	Anamnesis morbi / anamnesis vitae — onset, course, prior interventions.	0.95+
Objective	Physical-exam findings in structured form.	0.93+
Diagnosis (primary)	Clinical diagnosis as dictated by the clinician.	0.97+
ICD‑10 code	Mapped code, suggested — the clinician confirms.	1.000
Disease character	Acute, chronic, exacerbation, remission.	1.000
Secondary diagnoses	Comorbidities and concurrent conditions.	0.95+
Prescriptions	Drug, dose, route, frequency — validated against the drug reference.	0.92+
Recommendations	Follow-up, lifestyle, clarifying-question prompts for the clinician.	0.95+

SCRIBE automates data entry into the medical chart. Every record is reviewed and signed by the physician. Not certified as a clinical decision-support system.

Validation

Three validators per case. Two independent judges per scenario.

We only claim a pass when all three validators agree — twice, from judges running on different model vendors. The full methodology runs on every build.

Figure 3 Three-validator architecture. Two independent judge runs per scenario.

Format validator

Structural match. Are all nine fields present? Valid types? No schema violations? Deterministic, runs on every case.

schema pass

Keyword validator

Clinical terminology coverage checked against our 1,550-term medical dictionary. Flags drugs, ICD-10 codes, and disease descriptors that fall out of vocabulary.

1,550 terms · OOV flagging

Semantic validator · LLM judge

Domain-aware semantic equivalence against reference output. Every scenario runs twice, with LLM judges from two different vendors, to surface methodology bias.

two vendors · independent runs

One number is a marketing number. Two independent judges with an audit trail per case — that’s a benchmark.

Cost

Your electricity bill, not your per-transcript bill.

Scribe runs on hardware you already own or buy once. Marginal cost per visit approaches zero as soon as the GPUs are warm.

AROOGA SCRIBE on-prem	< $50 / mo	Electricity only — varies by DC tariff. Hardware amortised separately ($1.5K pilot capex).
Cloud scribe (generic SaaS)	$45–60K / mo	Per-transcript pricing at 14,500 visits. Not positioned for air-gapped regulated workflows.
DeepScribe / DAX (retail)	$75–150 / MD	Scales with headcount. No on-prem option; data leaves the hospital by design.

Compliance

HIPAA-ready architecture. Certification is a market-by-market track.

Scribe is deployment-ready for every regulated market we target. Certifications are separate work streams — announced on close, not before.

Market	Architecture	Certification status
United States	HIPAA-ready: edge PII masking, AES-256, audit trails, air-gap capable.	BAA track in progress
UAE / Gulf	On-prem deployment inside your perimeter, which addresses Federal Law No. 2/2019 residency requirements. Mapping to ADHICS controls and NABIDH / Malaffi connectivity in progress. English and Arabic clinical packs on the 2026 roadmap.	Architecture ready · language pack in progress
Turkey	KVKK-compatible deployment pattern; DPIA documentation available for the pilot.	Architecture ready · language pack in progress
European Union	GDPR-compatible deployment pattern — supported for partner deployments.	Supported (partner)

Edge PII maskingSensitive tokens masked client-side before any transmission.

AES-256 at rest and in flightPer-tenant keys. No cross-tenant data path.

Full audit trailEvery visit, edit, sign-off versioned with before-and-after diff.

Air-gap validatedClean-slate regression v0.4.5. Zero egress paths in air-gap mode.

The market

What we ship that the cloud scribes don't.

Feature comparison based on public product pages as of April 2026. We audit quarterly. Competitors here are excellent products — for a different buyer than you.

AROOGA Scribe compared to DeepScribe, Nuance DAX, Abridge, and Suki across six capability rows.
Capability	AROOGA Scribe	DeepScribe	Nuance DAX	Abridge	Suki
Air-gap deployment	✓	cloud	Azure only	cloud	cloud
Open-source core models	✓	—	—	—	—
Published two-judge benchmark	59 scenarios, two judges	white paper only	—	—	—
Edge PII masking pre-transmission	✓	partial	partial	—	—
Clinical-guideline validation on output	validator + dictionary	—	—	—	—
Sovereign deploy in regulated markets	✓	—	cloud-only	—	—

DeepScribe, DAX, Abridge, Suki are excellent cloud products. If your hospital’s lawyer or your country’s regulator will not let you stream patient audio to a US or EU cloud, they are not options for you. AROOGA SCRIBE is the on-prem option that passes the audit and ships numbers you can check.

Team

Built by people who came from clinical data.

Three of the engineers building Scribe came from a life-sciences data company, where they moved large volumes of clinical and preclinical data with automated conformance checks. Clinical data under regulatory constraint is not a new domain for this team — it is the one we came from.

Andrei

Founder & CEO

Owns product, go-to-market and the regulatory track. Built the company and the product from zero to a live clinical pilot without external capital.

Stanislav

Co-founder, international

Former deputy CTO of a major adtech platform. Solutions architect and technical product lead in enterprise software. Author of a professional course on ML and AI product management; mentor in technology commercialisation.

Dmitriy

Engineering direction

Seven years leading JavaScript and Java engineering departments at a life-sciences data company. Seven years before that as a technical architect.

Ivan

Backend and platform

Fifteen years in backend and infrastructure. Built clinical and preclinical data pipelines with automated conformance checks and prospective standardisation. AWS Certified Solutions Architect.

Timur

Business development

Commercial lead for public-sector and B2B channels. Built partner and dealer networks; grew average deal size fourfold in a public-sector channel.

Anastasia

Delivery

International delivery across teams of up to sixty people. Scrum and Kanban practice, English-speaking customer engagement.

Dr. Aleksey — clinical advisor Deputy chief medical officer of a state clinical polyclinic and a family physician by training; previously head of that polyclinic. Owns quality-of-care standards and guideline rollout inside the institution — the buyer of what we build, not only its user. Recognised by the national health ministry for contribution to healthcare, 2025.

Run a paid pilot in one clinic.

Bring three clinicians, your hardware or ours, your chart template. We wire Scribe into your dictation workflow and report structured accuracy weekly. Deployment inside your perimeter — no egress.

Request a pilot Architecture & compliance

3 clinicians · 4-week kick-off · data-processing agreement signed before go-live · weekly report on documentation completeness and clinician time

Questions

What hospitals ask on day one.

Does Scribe replace the physician?

No. Scribe automates data entry into the chart. Every record is reviewed and signed by the physician. Not certified as a clinical decision-support system.

What happens to the audio?

Recorded audio is processed in memory and is never written to disk as a recording, and never leaves the clinician's machine. One deliberate exception, documented rather than hidden: to survive a crash mid-consultation, the client holds a short rolling buffer of the current phrase on that same machine, encrypted with a per-session key and purged automatically. It is never transmitted. Structured outputs are stored in your EHR or in the Scribe local database per your retention policy.

Which EHR integrations work today?

The current pilot deployment uses an L0 clipboard adapter into the customer's EHR. The adapter pattern is EHR-agnostic; HL7 / FHIR, Epic and Cerner are on the integration roadmap — not shipped yet.

Which languages are in production today?

The current production deployment runs on a single-language clinical corpus (specific language disclosed under NDA for competitive reasons). English and Arabic clinical-language packs are on the 2026 roadmap and will publish on the same two-judge methodology before general availability.

What hardware do we need?

Pilot: 1× RTX 3090 (24 GB) + Ryzen 7 + 32 GB RAM ≈ $1.5K capex. Production at 3,000 clinicians: 2× A100 + 4× T4 + AMD EPYC 9274F ≈ $40–60K capex. We size per deployment.

Can we audit what executed behind each field?

Yes. Every inference runs locally on your infrastructure — no closed-model inference, no vendor API in the hot path. Under NDA we share the model manifest (weight hashes checksummed per release) so your InfoSec team can fingerprint exactly what ran, and verify reproducibility across deployments.

What if the clinician edits the output?

Every edit is logged with a before-and-after diff, timestamp, and user ID. Full audit trail per visit.

Is this HIPAA / GDPR / DHA certified?

The architecture is ready for all three. Certifications are market-by-market tracks we announce on close, not before. See Compliance above for current status.

Do you train on our data?

No. Customer audio and structured output are never used to train our models or anyone else's. This is architectural, not contractual.

What if the LLM hallucinates a prescription?

The LLM does not execute prescriptions. It populates a prescription field that the clinician reviews and signs. The validation layer compares extracted drugs to the 2,830-substance drug reference and flags anything out-of-vocabulary.

Clinical documentation, validated three ways.