TriageFlow AI

Inspiration

What it does## Inspiration

Emergency triage is a high-pressure workflow where clinicians must prioritize patients with incomplete information, varying measurement units, and limited time. We wanted to explore how AI could reduce intake friction and improve transparency without replacing clinical judgment.

What it does

TriageFlow AI is a voice-first clinical decision-support prototype for emergency triage.

Clinicians can begin with voice-assisted intake or structured manual entry. The system requests confirmation of required observations instead of assuming missing values, then applies local explainable EBM inference, deterministic safety guardrails, and calibrated uncertainty.

The workflow separates patient intake from the live triage queue to reduce cognitive load. Each completed assessment provides a local clinical rationale, safety status, and interpretable contributing factors.

Featherless AI can optionally generate a constrained supplementary explanation from approved local factor directions only. It cannot alter the local assessment.

How we built it

The backend uses Python and FastAPI with HL7 FHIR R5-oriented contracts. The local inference layer uses an Explainable Boosting Machine trained only on synthetic data, explicit lineage checks, deterministic data splits, and a versioned conformal calibration artifact.

The safety layer includes fail-closed validation, measurement confirmation, auditability, local fallback behavior, and deterministic guardrails that take precedence over model output.

The frontend is being implemented with Next.js, TypeScript, and a clinical design system focused on voice-first intake, progressive manual confirmation, an operational queue, configurable hospital themes, and high-legibility accessibility mode.

Challenges we ran into

The hardest challenge was defining safe boundaries around AI.

We had to ensure that:

• missing or ambiguous measurements do not become silent assumptions;
• local safety logic remains functional if external AI services fail;
• Featherless is useful in the demo without becoming part of the clinical decision path;
• model explanations remain interpretable and fail safely if malformed;
• voice intake remains clinician-confirmed rather than autonomous;
• synthetic evaluation is never presented as clinical validation.

We also found that perfect results on synthetic data can be misleading when generated categories correlate deterministically with labels. The current model is therefore a technical prototype, not evidence of clinical performance.

What we learned

Trustworthy clinical AI requires more than a prediction score. It requires data governance, uncertainty handling, explainability, safe degradation, interface design, and explicit human authority.

We also learned that voice-first interaction can reduce documentation burden only when every clinically relevant value remains visible and confirmable by a clinician.

What's next

Next steps are completing the voice-first frontend demo, deploying the prototype, recording a walkthrough, and pursuing approved evaluation on appropriately governed data.

Any future real-world clinical use would require clinical review, ethics and privacy approval, validation, and regulatory assessment.

Important limitation

TriageFlow AI is a hackathon prototype using synthetic data. It is not a medical device, does not provide autonomous diagnosis or treatment, and is not approved for clinical deployment.

Built With

• ai
• boosting
• conformal
• elevenlabs
• explainable
• fastapi
• featherless
• fhir
• hl7
• httpx
• interpretml
• machine
• next.js
• prediction
• pydantic
• pytest
• python
• react
• ruff
• typescript