HOPPR

HOPPR's foundation models are the engine of its core module, and much of what it sells stands without them. Grace, the multimodal imaging foundation model, and the narrative models (chest radiography in April 2026, 2D mammography in August 2026) are the models themselves.

The HOPPR AI Foundry around them adds accelerated computing, curated and labeled datasets, fine tuning tooling, a quality management system framework and traceable development records; Forward Deployed Services is a professional service; and Presto Agent, sold separately, is a model agnostic layer that delivers other models' results, open source, commercial or custom, into the radiologist's reporting environment. Compute, data, services and an integration layer carry value a customer can use with models that are not HOPPR's.

The Foundry takes no clinical action, and Presto Agent does, with oversight assumed rather than described. Presto is sold as AI draft reporting within existing radiology workflows: it places AI results, from models the practice chooses, into the radiologist's reporting application (a radiologist on the product page says it runs inside PowerScribe) without changes to PACS or reporting systems, and radiologists have described it inserting measurements they would otherwise dictate.

The radiologist signs the report, but nothing published states what Presto may insert without confirmation, whether inserted content is marked as model generated at the point of signing, or whether the signed report records which model produced which content. A number placed in a template looks the same whether right or wrong, and checking it means remeasuring, which is the work the feature removes. For the Foundry, nothing addresses what oversight design downstream builders are expected to include.

What produces the output is named, characterized and versioned. Grace is identified as HOPPR's multimodal foundation model for X-ray, CT, MRI and echocardiography, supporting image to image and text to image learning, and its training corpus is characterized as more than a petabyte of permission based, anonymized imaging studies paired with their reports, across 2D, 3D and longitudinal series; the company describes de identification at ingestion and curated, labeled data with known provenance.

The narrative models ship with training data traceability records and version locking so developers can reproduce results, and the Foundry keeps traceable records of dataset usage, training configurations and every model update. Third party models on the platform are named rather than white labeled (NVIDIA's NV-Reason and NV-Generate since March 2026). Still unpublished: the geographic, scanner and demographic composition of the corpus.

Substantial partial disclosure. The model layer is HOPPR's own Grace and narrative models, with NVIDIA's open models NV-Reason and NV-Generate named where they run on the Foundry, and the Foundry is stated to run on Amazon Web Services. No subprocessor list is published, the AWS region is not stated, and the privacy policy (updated 21 August 2025) refers to service providers under written contracts without naming them beyond Google Analytics.

The terms that govern customer data once it arrives are the gap: whether data brought into the Foundry for fine tuning is isolated from the base models and from other customers, and what Presto Agent passes to the third party models a practice chooses to run.

Evidence appropriate to infrastructure is thin. The chest radiography narrative model is described as evaluated against internal benchmarks with no published results, comparison or external validation, and no downstream clinical application built on Grace is named with outcome data. For Presto Agent, radiologists at named practices (Evansville Radiology, Radiology Physicians and Radiology Associates of South Florida) describe daily use without figures.

Partners and investors (DeepHealth, TestDynamics, Health2047, the American Medical Association venture studio) are relationships, not performance evidence. For a base model others build regulated products on, published benchmark performance is the reasonable expectation.

A categorical commitment on the training corpus is published, and the handling of customer data is not. HOPPR describes Grace's training corpus as permission based and anonymized, with de identification at ingestion, and the Foundry is covered by SOC 2 Type II attestation and HITRUST e1 certification.

What is not stated is the de identification standard, the permission mechanism, retention for customer data (the privacy policy keeps information for as long as necessary), and whether data brought into the Foundry for fine tuning can reach the base models. Presto Agent processes live report content in the practice's reporting environment, and nothing describes what it retains. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

A privacy document reaches the service and no business associate status is stated. HOPPR's privacy policy (IntellixAI, Inc., doing business as HOPPR; updated 21 August 2025) covers website visitors and users of its services, including the AI Foundry and Presto Agent, which processes medical imaging data, and describes the company as a data controller; it contains no HIPAA or business associate statement, and the security page directs buyers to sales for agreement terms.

The Foundry holds HITRUST e1 certification and SOC 2 Type I and Type II attestations, which are security assessments rather than a data protection position. The company runs three flows with different postures: an anonymized base corpus, customer data brought in for fine tuning, and live report content in Presto.

Recognized attestations are named for one of the two products. HOPPR states SOC 2 Type I and Type II attestations (Thoropass named on the Type I badge) and HITRUST e1 certification for the AI Foundry after an independent assessment, on Amazon Web Services, with access management and least privilege, continuous monitoring and logging, incident response and governance described. Naming the e1 tier, the entry level of HITRUST's three, is precise.

No trust center, report download, penetration test disclosure or attestation period is published, and detailed documentation is routed through sales. The certifications are stated for the Foundry; Presto Agent, a separate product running inside practices' reporting environments, is not placed inside the certified boundary.

No clearance is claimed, and none is apparent for the Foundry, which supplies models and development infrastructure; regulatory responsibility for a clinical product built on them sits with the developer who submits it, and using a cleared component does not confer clearance. The company describes a quality management system framework, regulatory alignment and traceable development records meant to support compliant medical AI development, without naming the regime it aligns to.

Presto Agent is sold as AI draft reporting inside the radiologist's reporting environment and states no regulatory position. It does not supply the analysis, but it is the layer through which other manufacturers' outputs, cleared or not, reach the signed report, and it can change the conditions under which a cleared product is used.

No subgroup analysis, bias assessment, fairness evaluation or model card describing the training distribution is published. The training corpus is quantified (more than a petabyte of permission based, anonymized studies with reports across modalities) but not composed: nothing states the geographic origin of the studies, the scanner and field strength mix, the academic and community mix, or patient demographics.

That matters more for a base model than for an application, because a developer building a device on Grace inherits its distribution and must characterize its own product's performance across subgroups, and only the model provider can describe what it inherited. Presto Agent places validation and monitoring of the models a practice chooses with the practice, and nothing describes what a practice must establish before wiring a model into live reporting.

Mechanisms exist that let a downstream developer trace and reproduce what a model did, and nothing stands behind the output. The narrative models ship with training data traceability records and version locking, and the Foundry records dataset usage, training configurations and every model update, which lets a developer building a regulated product on HOPPR's models reconstruct and defend what it built on.

No error rate with its method and denominator is published for Grace or the narrative models, which are described as evaluated against internal benchmarks, and no accuracy commitment, service level, indemnity or remediation obligation is published, toward developers or toward practices using Presto Agent.

Integration is the substance of Presto Agent. It places AI results into the radiologist's existing reporting application without changes to PACS, reporting platforms or templates, and a radiologist quoted on the product page says it runs inside PowerScribe; practices named on the page include Evansville Radiology, Radiology Physicians and Radiology Associates of South Florida. That is a named reporting system receiving one directional output. No compatibility list, integration standard or EHR connection is published.

The Foundry runs on Amazon Web Services, with access management, least privilege and isolation named as design properties; no region, residency commitment or tenancy model is published, and nothing says where fine tuning compute runs. Presto Agent is installed into the practice's existing reporting environment without changes to PACS or reporting systems, and where it processes report content is not stated.

No price is published. The Foundry is sold as platform access to developers, AI companies and imaging software vendors through an access request, with Forward Deployed Services as a separate professional services engagement, and Presto Agent is sold separately to radiology practices.

Scope is clear on modality and buyer. Grace spans X-ray, CT, MRI and echocardiography across 2D, 3D and longitudinal series, and narrative models cover chest radiography (April 2026) and 2D mammography (August 2026). The Foundry is sold to developers, AI companies and imaging software vendors, and Presto Agent to radiology practices, with named practices using it. No subspecialty scoping is stated, and the modality list describes what the model was trained across rather than what it was evaluated on.