Cortex
Fine-tunes clinical reasoning models strictly inside the hospital's private compliance boundary, guaranteeing zero PHI training data egress.
Engineering highlights
The core architectural breakthroughs.
High-scale engineering demand distilled into four verified production milestones.
Strict in-tenancy Ray/PyTorch training inside private hospital VPC
Trained on 4.8M de-identified patient notes with differential privacy
Cycle time reduced from 18 days via LoRA/QLoRA parameter efficiency
Dynamic GPU autoscaling eliminates overnight idle cluster costs
Executive overview
The engineering challenge & solution.
Hospital physicians required fine-tuned clinical diagnostic assistance trained on 4.8 million historic patient charts, pathology notes, and radiology reports. However, HIPAA regulations, institutional review boards, and privacy laws strictly prohibited sending patient data to public cloud APIs.
We built an on-premise private VPC training and inference fabric using Kubernetes, Ray, and PyTorch. We implemented an automated NLP de-identification pipeline that strips 18 HIPAA identifiers before data curation, combined with a LoRA parameter-efficient fine-tuning harness operating across a private cluster of NVIDIA GPUs.
Engineering governance
The three non-negotiables.
These are the architectural constraints that shaped every boundary — no trade-offs or compromises permitted.
Zero PHI Data Egress
Absolute physical and cryptographic isolation: no training token or diagnostic query leaves the private hospital VPC.
Full SHA-256 Training Lineage
Every model checkpoint must be cryptographically linked to the exact de-identified training dataset and hyperparameter run.
Continuous Clinical Eval Gates
Automated diagnostic benchmarks evaluating clinical accuracy against board-certified physician test panels.
Deterministic Automated De-Identification and Differential Privacy Auditing at Scale
Clinical notes contain unstructured medical text where patient names and family relationships are buried inside narrative sentences. Standard regex redaction missed 7% of identifiers, while heavy models were too slow for 4.8M records.
We developed a specialized hybrid de-identification pipeline pairing spaCy biomedical entity recognition with a validated transformer model. We added a differential privacy Laplace noise mechanism to the gradient updates during LoRA training, mathematically proving that patient records cannot be reconstructed through model weight extraction.
Request lifecycle
How data moves through Cortex.
Production verification
Measured outcomes in production.
Every metric below was captured under real production traffic and audited against historical baseline data.
Operational resilience
Automated safety guardrails.
Systems fail in production. Here is how Cortex survives network partitions, upstream timeouts, and anomalous inputs without human intervention.
TPM Hardware Key Enclaves
Model weights and patient datasets encrypted at rest using physical hardware keys.
Strict SCIM Role-Based Access
Physician roles govern diagnostic data access with full audit logging.
Automated Clinical Eval Gates
New weights blocked from deployment if benchmark recall drops below 92%.
Next step
Building a system with this level of demand?
Bring us the constraint that keeps your engineering leadership up at night — the latency ceiling, the compliance perimeter, or the unyielding reliability requirement.