Platform
The model is one component. iTech builds the rest: criteria engines, clinical context, quality control, provenance, orchestration and integration, arranged so a qualified professional stays in charge.
Four layers
The engine
Deterministic criteria, structured clinical reasoning, recommendations, QA, provenance and literature linkage. This is the layer the other layers are built on.
Proof in real workflow
A complete radiologist-facing implementation of the stack: report construction, decision support, classifications, QA, recommendations, citations, point-of-care learning and group governance.
Callable from anything
The same intelligence exposed to PACS, RIS, EHR, imaging-AI, reporting and foundation-model companies, and to autonomous agents, over a versioned API.
Adjacent research platform
Persistent semantic memory, multimodal identity, real-time conversation and consent-governed biometrics: how intelligence remembers, recognizes and persists. Not a medical product.
Architecture
General-purpose AI can generate extraordinary language. Clinical intelligence also requires maintained knowledge, deterministic rules, clinical context, provenance, quality control and a qualified professional at the end of the chain.
Dictated findings, report text, images or structured fields.
Free text becomes canonical findings: size, composition, laterality, with honest confidence.
Age, risk, screening status and priors that change which standard applies.
A validated criteria pack decides the category. Same input, same answer.
A language model writes and interprets, inside the verdicts it was given.
Mechanical checks: laterality conflicts, recommendation concordance, and prose that strays outside the verdicts.
Criteria pack, version, validation status and trace id travel with the answer.
A qualified professional reviews, edits and signs. The final authority stays human.
iTech builds the architecture around the model.
Technology
Clinical rules should not be reinvented by a generative model on every request. Criteria packs decide categories the same way every time.
Models are powerful for interpretation, synthesis and communication. They write inside the verdicts, and a mechanical guard checks that they did.
Expose which criteria, which version and which evidence produced an answer, so it can be audited after the fact.
Model capabilities change quickly. The architecture routes across vendors, falls back when one fails, and records which model actually ran.
The qualified professional remains the final authority. Software drafts, checks and cites; it does not sign.
Beyond generation
A model turns pixels and words into representations. Turning a representation into a clinical decision takes criteria, a version, evidence and a record of how the answer was reached.
Domain intelligence
Clinical decisions should expose the criteria, version, evidence and reasoning context that produced them. Not because AI should be trusted blindly. Because it shouldn't have to be.
Available via API: Callable today through RadVector API. Versioned, validated criteria packs answer through the API today.
Coming soon: In development. Not yet available. A maintained corpus of published guidance and newer evidence is being wired in.
Decision
Input: 7 mm solid nodule, CT thorax, risk high
RecommendationCT at 6-12 months, then CT at 18-24 months.
From a real RadVector API response. Clinical decision support for qualified professionals.
Validated · may drive output
Every verdict names who validated its pack and when.
Draft · answers as advisory
Draft packs are marked not usable for output. A deferral always says why.
Model orchestration
Clinical AI should not depend on a single model vendor. iTech routes different workloads across multiple commercial and open-weight models, while clinical governance stays above the model layer.
Construction. Builds the report through FINDINGS from the radiologist's own words. Findings authority is enforced mechanically downstream.
Reasoning. Drafts IMPRESSION and RECOMMENDATIONS around the deterministic verdicts instead of recalling a guideline from memory.
QA. Can run on a different model than the one that drafted, so a second system checks the first.
Vision. Reads worksheets, tech sheets and images, behind an OCR-based PHI gate.
References. Checks cited claims against a registry of published sources and shows what was consulted.
Every report records the model that actually ran, with its tokens and cost. Organizations choose their vendors and models, including their own keys on enterprise plans.
Ecosystem
The same clinical engine reaches radiologists through RadVector Reporting and software, AI vendors and agents through RadVector API. Connections are labeled by what exists today.
Select a surface to see what exists today.
Radiologist via RadVector Reporting Live: In production in a shipping product today.
Radiologists use the intelligence layer inside RadVector Reporting: report construction from their own findings, criteria, QA, recommendations and references, with the radiologist signing every report.
EHR / RIS Integration-ready: Built and available for integration projects.
RIS REST and FHIR R4 clients read worklists, and an HMAC-signed HL7 v2 listener accepts orders (ORM), for integration projects. A browser connector links RIS worklists today.
AI agent Available via API: Callable today through RadVector API.
An AI agent can call RadVector API as a tool today: a finding goes in, a versioned verdict with provenance comes out, and a deferral says why when the criteria do not apply.
PACS / VNA Coming soon: In development. Not yet available.
PACS and DICOMweb connectivity is in development and not yet available. The design goal is to add intelligence alongside the viewer and archive, not replace them.
Modality Direction: Architectural direction. Not a current product capability.
Architectural direction: combining acquisition context with criteria and clinical history downstream. No modality integration exists today.
AI vendor Available via API: Callable today through RadVector API.
Imaging-AI vendors can pass a detection to RadVector API and receive the criteria-based category and recommendation that follows from it.
Software Available via API: Callable today through RadVector API.
Reporting, PACS, RIS and EHR software can call the same deterministic engine that powers RadVector Reporting, over a versioned REST API.
Research Available via API: Callable today through RadVector API.
Research teams can apply versioned, validated criteria consistently across de-identified datasets, with a trace id on every verdict.
Walk through RadVector Reporting, RadVector API and the integration surfaces with the team that built them.