Inspiration
Healthcare AI is entering a dangerous phase.
LLMs are now capable of generating prescriptions, discharge recommendations, diagnostic suggestions, and entire clinical workflows. But there’s a fundamental problem:
Most healthcare AI systems focus on generating decisions.
Very few focus on validating whether those decisions are actually safe before execution.
And in healthcare, hallucinations are not just inconvenient — they can directly harm patients.
I kept asking myself a simple question:
What happens when an AI healthcare agent confidently makes the wrong clinical decision?
For example:
- prescribing Penicillin to a patient with a documented Penicillin allergy
- discharging a clinically unstable patient
- duplicating medications or procedures because of fragmented patient context
Today, many healthcare AI workflows still rely heavily on probabilistic reasoning without deterministic runtime safeguards.
That felt fundamentally unsafe.
So instead of building another healthcare chatbot, I decided to build the missing infrastructure layer behind healthcare AI systems.
That became CROSS.
What CROSS Is
CROSS is a deterministic runtime clinical safety infrastructure layer for healthcare AI systems.
Instead of generating medical decisions, CROSS validates whether AI-generated clinical actions are safe to execute before they enter downstream workflows.
The system acts like a runtime safety gateway between:
- AI-generated clinical actions
- and real-world execution
CROSS transforms fragmented FHIR patient records into structured runtime context, performs pre-execution safety simulation, and deterministically enforces GO / NO_GO runtime safety decisions.
This allows healthcare AI agents to consult CROSS before:
- prescribing medications
- ordering diagnostic tests
- executing discharge actions
- performing workflow automation
Core Runtime Workflows
CROSS currently exposes three core runtime workflows:
Runtime Context Builder
Normalizes fragmented FHIR patient data into reusable runtime clinical state.
Pre-Execution Safety Simulation
Predicts runtime safety outcomes before execution occurs.
Deterministic Runtime Enforcement
Validates proposed clinical actions against structured patient context and blocks unsafe actions using explainable GO / NO_GO decisions.
Current runtime safety rules include:
- medication allergy conflicts
- duplicate medications
- duplicate procedures
- unsafe discharge conditions
How I Built It
CROSS was built as an MCP-native healthcare infrastructure system.
The backend runtime engine was implemented in Python using FastMCP with deterministic validation workflows operating on structured FHIR-style patient context.
I built:
- a FHIR runtime context normalization layer
- a stateless clinical action simulation engine
- deterministic runtime validation pipelines
- explainable safety enforcement outputs
- MCP tools for simulation, validation, and context construction
- a lightweight workflow orchestration agent connected to the CROSS runtime layer
The project operates on synthetic FHIR patient records containing:
AllergyIntoleranceMedicationRequestObservationProcedureCondition
Each proposed clinical action is evaluated against this runtime context before execution occurs.
Challenges I Faced
One of the biggest challenges was resisting the temptation to build a “smart healthcare chatbot.”
I intentionally designed CROSS as infrastructure instead of a medical reasoning assistant.
That meant prioritizing:
- deterministic enforcement
- explainability
- runtime safety
- operational behavior
- structured context grounding
Another challenge was designing workflows that felt operational rather than conversational.
I wanted CROSS to behave like a runtime clinical safety layer that healthcare systems could realistically integrate into orchestration pipelines.
I also spent significant time designing MCP-native workflows that could cleanly expose runtime simulation and validation as reusable infrastructure capabilities.
What I Learned
Building CROSS changed how I think about healthcare AI.
I realized the future of healthcare AI is not just about making models more capable — it’s about building deterministic safety infrastructure around them.
In healthcare, reliability matters as much as intelligence.
I also learned that MCP-native architectures are extremely powerful for building reusable AI infrastructure layers that multiple healthcare agents and workflows can consult before execution.
Future Vision
My long-term vision for CROSS is to become a reusable runtime clinical safety layer for healthcare AI ecosystems.
Future directions include:
- longitudinal patient-state tracking
- richer medication interaction validation
- runtime audit trails and explainability traces
- orchestration-aware safety policies
- broader healthcare workflow integrations
- deeper FHIR interoperability
As healthcare AI systems become more autonomous, I believe deterministic runtime safety infrastructure will become essential.
CROSS is my attempt at building that foundation.
Log in or sign up for Devpost to join the conversation.