From research code to a clearable, deployable product

Stages 1–3 · Retrieval, Ingest, De-identify

Data aggregation & de-identification

Stand CuratAI up inside your firewall. Connectors get configured against your PACS, EHR, pathology system, and research data sources — REDCap, Excel and CSV exports, registry pulls; the local LLM gets tuned to your registry's fields, and the de-identification rules to your data's quirks. By the end of the engagement, your team owns a working pipeline that turns raw clinical data into de-identified, structured research cohorts.

• PACS / EHR / pathology connector setup
• Research data ingest from REDCap, Excel/CSV, registry exports
• Local LLM tuning for your target registry
• De-identification rule customization across 5 PHI vectors
• IRB-compatible audit + sharing configuration

Stage 4 · Annotate

Annotation & cohort building

Annotation is the bottleneck on every clinical-AI project. Multi-user projects get set up in CuratAI, the team gets trained on AI-assisted annotation, and the review workflow ships with QA gates baked in. When the workload outruns the team, our clinically-trained annotators step in as a managed service. When the question calls for data outside your walls, we tap an external network of imaging-data providers to assemble the cohort.

• Annotation project + protocol design
• AI-assisted annotation workflow training (prediction + interpolation)
• Multi-user roles, reviewer queues, audit trails
• Managed annotation by our clinical team (optional)
• External imaging-data sourcing when the cohort needs to come from outside

Stage 5 · AI Plugins

Custom AI development

Full-spectrum clinical-AI engineering, augmented by the CuratAI platform that handles data plumbing, audit, and on-premise deployment. The team has shipped FDA-cleared AI on this stack twice; the same team and software come to your model. Optimization, benchmarking, and gap analysis are available as standalone engagements when you already have a model in hand.

• Custom model training on your de-identified data
• Retraining of FDA-cleared products (INTContour, OncoAI) on your protocols
• Performance benchmarking + gap analysis on existing models
• Algorithm optimization for accuracy, latency, and clinical workflow fit
• Packaging as a CuratAI plugin or standalone on-premise deployment

See: Mayo Arizona INTContour retraining →

Stage 6 · Collaborate · into the clinic

Regulatory & 510(k)

From research code to a clearable product. We've shipped two FDA 510(k)-cleared products (INTContour K212274, INTDose K213137) and run the full lifecycle: pre-registered validation study design, predicate selection, pre-submission consultations with FDA, submission preparation, deficiency-letter response, post-market surveillance. The validation tooling we built for our own clearances comes with the engagement: action-logged annotation, automated metric computation against the reference standard, statistical analysis for substantial equivalence. It produces an FDA-conforming evaluation report straight out of the run.

• Pre-registered validation study design + statistical analysis plan
• Pre-submission (Q-Sub) consultations with FDA
• Predicate selection & substantial-equivalence argument
• Risk assessment and cybersecurity content for the submission dossier
• Auto-generated, FDA-conforming evaluation reports from validation runs
• 510(k) submission preparation and deficiency-letter response
• Post-market surveillance + change-control process