Ingest any document. Ask it anything.
Cortex reads PDFs, scans, faxes, and handwritten notes, then answers questions about them in plain English. Every answer comes back with a page-level citation in under a second, and the data never leaves your network.
Built for the environments where cloud AI is off the table: HIPAA, SOC 2, 21 CFR Part 11, FedRAMP, and fully air-gapped.
Bring us your hardest documents
Query response
Documents, single GPU
Compute per page
Documents sent off-site
Every regulated business has the same locked vault.
Thousands of PDFs, scans, faxes, and handwritten notes hold the answers to the most expensive questions in the business. Was medical necessity documented? Which contracts auto-renew in 90 days? Does this loan file match three years of W-2s?
The answers exist. Today, the only key is a person who reads.
Reading is the most expensive query engine ever built.
4.7B
Claims processed by health payers each year at $12 to $19 each, largely by hand.
60–70%
Of a law firm associate’s time goes to document review. One data room: 6 associates, 3 weeks, 15,000 contracts.
80%
Of an intelligence analyst’s time goes to finding information, not analyzing it.
$480K
Locked in for another 24 months by one missed non-renewal notice. One mistyped customs code costs up to $50,000 and 3 to 5 days.
Keyword search does not fix this.
It matches words, not meaning. It misses the clause worded differently, and the “myocardial infarction” you searched as “heart attack.”
Cloud AI could fix it. Compliance says no.
HIPAA, SOC 2, 21 CFR Part 11, and FedRAMP take that option off the table for exactly the documents that matter most. So the reading continues.
A document intelligence platform that runs entirely on your infrastructure.
Upload any PDF or scan: typed, faxed, stamped, or handwritten. GPU-powered OCR reads it, semantic indexing understands it, and from that moment anyone can ask questions in plain English.
Every answer comes back three ways
Cited
Document name, page number, and confidence score on every claim. One click opens the original, scrolled to the exact page. Staff verify instead of trusting a black box.
Structured
Define the JSON schema you want and answers flow straight into your loan origination system, claims platform, TMS, or EHR. No re-keying.
Recorded
A full audit trail of every query and answer. For a compliance officer facing a surveyor or regulator, the audit trail is the product.
Time-range filtering
Lookback questions answer directly: “what medication changes in the last 14 days?”
Per-client isolation
Every tenant’s documents stay hard-partitioned from every other tenant’s.
The category splits in two, and neither half solves the problem.
Legacy document processing suites read documents but cannot answer questions: you have to know the field before you can extract it. Modern document-AI APIs parse well but run in the cloud, a non-starter for regulated data. Cortex combines GPU OCR, semantic question answering, schema-flexible extraction, page-level citations, an audit trail, and on-premise deployment in one system.
| Approach | Answers open questions | Runs inside your network | Cited and audit-logged |
|---|---|---|---|
|
Legacy IDP suites Template and field extraction. You must define the field first. |
No | Usually | Partial |
|
Cloud document-AI APIs Strong parsing, delivered as a hosted service. |
Partial | No | Varies |
|
Cortex Semantic Q&A plus schema-flexible extraction on your hardware. |
Yes | Yes | Yes |
The numbers
Accuracy
In early clinical validation, medication extraction went from 26% to 93% and allergy extraction from 15% to 100% against baseline OCR pipelines. Current models score higher still.
Accuracy is never a trust exercise here. Every answer cites its page, so you measure it on your own documents in week one, not on our benchmark.
Throughput
315 documents per hour on a single GPU. 2 to 5 seconds per page. Queries return in under a second.
Cost
Roughly $0.001 per page in compute, 50 to 200x below per-page document-AI API pricing.
Staffing
Exception-only workflow. More than 80% of documents process straight through, and staff only see the 10 to 20% that needs human judgment.
The Q&A engine is a foundation, not a feature.
Higher-level products are compositions of queries, not new builds. Each of these is configuration on top of the same engine.
Compliance Monitor
Run a checklist against every document nightly and surface what failed.
Event Detector
Watch for the conditions that matter and flag them as documents arrive.
Trend Analyzer
Track how answers move across a corpus over time.
Risk Scorer
“Score this loan file on 12 risk dimensions,” with the page behind every score.
Workflow Automator
Route the exceptions to the right desk and let the rest process straight through.
Regulatory Tracker
“Which of our SOPs does this new FDA guidance affect?”
Any industry where the documents are the bottleneck.
Our healthcare depth is the proof, and the pattern holds anywhere reading gates the decision.
| Industry | What changes |
|---|---|
| Healthcare | Payer prior authorization, SNF and PDPM reimbursement accuracy, clinical trial TMF search, coding and denial appeals, radiology and pathology extraction, CMS survey readiness in minutes. See our healthcare practice. |
| Financial Services | Loan underwriting and KYC/AML. Banks spend $150M a year on KYC. Income verification drops from 3 to 5 days to seconds. |
| Legal | M&A due diligence and contract review. 3 weeks of associate review becomes 2 days of verification. |
| Insurance | Claims processing and fraud detection. Patterns surface across 50,000 historical claims no adjuster can remember. |
| Logistics | Trade compliance across the 50 documents and 30 parties behind every international shipment. See our logistics work. |
| Manufacturing | SOP version control and FDA audit prep. 2 weeks of assembly becomes a set of queries. |
| Government & Defense | FOIA backlogs and intelligence analysis, fully air-gapped. See our defense work. |
| Energy & Utilities | NERC, FERC, and EPA compliance, where violations run $1M per day. See our energy work. |
| Real Estate | Lease abstraction. 3 paralegals for 4 weeks becomes a query, amendments included. |
An accelerator, not a SaaS subscription.
Cortex is one of the Ventures accelerators we deploy inside client engagements. It runs on your hardware or a dedicated GPU cloud instance, it is configured against your documents and your workflows, and what gets built on top of it is your IP.
Department pilots start around $3K per month. Enterprise and air-gapped government deployments scale from there.
Deployment
Your data center, your VPC, or a dedicated GPU instance. Air-gapped installations supported.
Integration
Schema-defined output lands in the systems you already run: LOS, claims platform, TMS, EHR.
Ownership
Client IP from inception, on the same Hybrid BOT terms as the rest of Ventures.
The pilot is simple. Bring us a box of your hardest documents.
Faxed, stamped, handwritten, whatever your team dreads most. Then ask them anything, and check every answer against the page it came from.