Criteria extraction
NeMo NLP reads protocol PDFs and produces structured inclusion/exclusion criteria — diagnosis, labs, timing, prior treatment.
Product · Clinical trial matching
Orbis —Every patient finds their trial.
Orbis uses NeMo NLP to extract trial criteria, then runs RAPIDS GPU-accelerated semantic matching over your entire specialty patient population — identifying eligible patients in under 2 minutes and surfacing ranked matches in the clinician's Scribara workflow.
Layer 4 — clinical research access built into the specialty encounter workflow
NeMo RetrieverRAPIDSClinicalTrials.govNVIDIA InceptionHIPAAIRB-ready
NeMo criteria extractprotocol NLP
RAPIDS patient index100k+
GPU similarity match<2 min
Encounter contextScribara Labs
The problem
Why Orbis exists
Fewer than 5% of eligible patients are ever informed of the trials they qualify for — not because trials are rare but because matching at scale is prohibitively labor-intensive. 80% of trial delays are caused by slow enrollment. Specialty practices have the exact patient populations sponsors need but lack the tools to surface them.
- NeMo NLP extracts structured inclusion/exclusion criteria from protocol PDFs
- RAPIDS GPU vector similarity: 100k+ patients × 200+ active trials in under 2 minutes
- Match cards enriched with last Scribara encounter — notes, codes, imaging, medications
- Continuous nightly re-screening as registries update and new encounters are added
01 / Extract
Criteria, structured from protocol.
NeMo NLP reads ClinicalTrials.gov protocol PDFs — extracting inclusion/exclusion criteria as structured entities: diagnosis codes, lab thresholds, age ranges, prior treatments, imaging criteria.
02 / Index
Your panel, vector-encoded.
Patient records are embedded by NeMo and indexed in a RAPIDS GPU vector store — 100k+ patients queryable in seconds, updated nightly as new Scribara encounters are completed.
03 / Match
100k patients, 200 trials, 2 minutes.
RAPIDS cuML GPU-accelerated similarity scoring runs the full panel against all active trials, returning confidence-ranked match lists — a workload that takes 90 minutes on CPU.
04 / Notify
The right patient, at the right visit.
Match cards appear in the clinician's Scribara workflow with encounter context and a NeMo-drafted "why this patient qualifies" summary. Approved matches trigger consent workflow via Authra. [ASPIRATIONAL]
Capabilities
What Orbis does
GPU matching
RAPIDS vector similarity: 100k+ patients × 200+ trials in under 2 minutes. CPU-only would take 45–90 minutes per run.
Encounter-enriched cards
Every match includes the patient's last Scribara notes, codes, and imaging findings — and a NeMo-drafted eligibility summary. [ASPIRATIONAL]
Continuous screening
Nightly re-screening as trial registries update and new encounters are added — no patient is missed.
Consent workflow
Clinician-approved matches trigger consent documentation and enrollment steps via Authra. [ASPIRATIONAL]
Sponsor pipeline
Pharma sponsors access de-identified aggregate eligibility data for site identification — a B2B revenue stream. [ASPIRATIONAL]
GPU-essential
Why accelerated compute is required
Matching 100k patient embeddings against 200 trial criteria vectors — 20 million similarity calculations — must complete in under 2 minutes for practical nightly re-screening. RAPIDS GPU reduces 90-minute CPU runs to under 2 minutes. NeMo criteria extraction from 30-page protocol PDFs runs in under 5 seconds on GPU vs. 90 seconds on CPU.
- RAPIDS cuML vector similarity: 20M calculations in <2 min
- NeMo Retriever: cross-encoder reranking at TensorRT speed
- NeMo NLP criteria extraction: <5 s/protocol on GPU vs. 90 s on CPU
- NIM on NVIDIA AI Enterprise for on-prem health-system deploy [ASPIRATIONAL]
RAPIDS cuMLvector similarity
NeMo Retrieverreranking
TensorRT criteria<5 s/protocol
Triton servingmulti-tenant
Impact
What it moves
0
Full panel × all trials on GPU
0
Active trials matched per run
0
Current enrollment rate for eligible patients
0
Clinical trial recruitment market (US, 2026)
The ecosystem
Works with the rest of Scribara
Orbis uses encounter data from the full platform and routes enrolled patients back into monitoring and analytics.
Answers
Frequently asked
ClinicalTrials.gov API is the primary source; pharma sponsors can also push protocols directly via our intake API. IRB and IRB-exemption guidance [UNKNOWN] — confirm with your institutional compliance office.
Patient embeddings are computed from de-identified Scribara encounter outputs. Patient identity is never shared with sponsors; only de-identified aggregate eligibility data is made available for site identification.
Yes — the sponsor pipeline provides de-identified aggregate eligibility counts (not individual patients) for site identification. Full sponsor product is [ASPIRATIONAL]; contact us for early-access terms.
Connect your patients to the trials that need them.
See Orbis screen a specialty panel against active cardiology trials in a live demo.