Solace — AI-native ED intake & triage

AI-native patient intake and clinical triage for emergency departments — multilingual voice intake and explainable ESI, on Amazon DynamoDB, deployed on Vercel.

Comment

Inspiration

Emergency departments lose time on the front end of every single encounter. Registration pulls nurses off clinical work, language barriers cause measurable under-triage for non-English speakers (1 in 5 ED visits in major metros involves limited English proficiency), and clinicians flip through paper forms or multiple EHR screens before they even see the patient. We wanted to erase the gap between a patient walking through the door and a clinician knowing exactly what they need — without asking the patient to download anything.

What it does

A patient scans a QR code in the waiting room and speaks their symptoms in any of 20+ languages — no app, no account. Within ~7 seconds they hear a warm, spoken explanation of their triage level and what to expect. They can optionally photograph an injury or insurance card, and tap a button if their pain worsens — which instantly alarms the clinician board.

On the other side, the clinician already has a full AI pre-brief before the patient is roomed: a provisional ESI acuity level, a SHAP explanation of why the model scored it that way, a conformal prediction interval that flags when the model is uncertain, matched EHR data (allergies, meds, conditions, prior visits), and a one-click AI scribe draft. Bedside vitals refine the ESI in real time. Patients who can't use the QR can call a Twilio voice agent that handles triage, scheduling, and 911 escalation.

How we built it

  • Frontend: Vite + React 18 + TypeScript + Tailwind + Framer Motion. Patient intake, a draggable split-screen clinician "showcase," WebGL hero. Strong accessibility (hundreds of focus-visible targets, aria-live regions, RTL support).
  • Backend: FastAPI + Mangum on AWS Lambda (container, arm64, Python 3.12) behind API Gateway + CloudFront + WAFv2.
  • ML (the centerpiece): a 4-model stacked ensemble (LightGBM + XGBoost + CatBoost + MLP), 5-fold CV, logistic-regression meta-learner, ordinal threshold optimization, split-conformal calibration for 90% coverage, and per-patient SHAP via pred_contrib. A deterministic clinical safety floor can only raise acuity, never lower it. The trained ensemble is live in production (the /health endpoint proves trained_ensemble).
  • AI services: Claude (triage narrative, scribe, pre-brief, vision OCR), AWS Transcribe (STT), AWS Polly (TTS). A content guard redacts 18 HIPAA Safe-Harbor identifiers before any third-party LLM call.
  • Integrations: SMART-on-FHIR PKCE OAuth (Epic/Cerner/Athena), Twilio voice/SMS.
  • Security/compliance: one customer-managed KMS key across DynamoDB/S3/Secrets Manager, TLS 1.2+, dual-write audit logging (90-day hot + 6-year cold), brute-force lockout, identity-keyed rate limiting, prompt-injection scanning. All provisioned by 28 idempotent scripts. See SECURITY.md and docs/HIPAA_COMPLIANCE_DUE_DILIGENCE.md.

Challenges we ran into

  • Keeping train/serve feature engineering perfectly in sync between the training notebook and the Lambda inference path.
  • Making a synthetic-data-trained model clinically safe — solved with a deterministic safety floor rather than trusting raw probabilities.
  • Real serverless HIPAA posture (CMK everywhere, BAA-covered services, audit retention) without a static server.

Accomplishments we're proud of

  • Real explainability (SHAP + conformal intervals), not a static chart — the clinician sees why and how confident.
  • A genuinely production-grade security/compliance story implemented in code, not slideware.
  • A patient experience that works on any phone in 20+ languages with zero install.

What we learned

Honest ML in healthcare means quantifying uncertainty and building deterministic guardrails around a probabilistic model — and that the hardest part of "AI for clinicians" is explainability and trust, not raw accuracy.

What's next

Real-world clinical validation study, EHR write-back (currently read-focused), per-tenant multi-hospital provisioning, and a formal SaMD/FDA pathway analysis for the acuity model (clinician-in-the-loop advisory posture). See docs/roadmap-50-features.md.

Built With

  • anthropic-claude
  • api-gateway
  • aws-bedrock
  • aws-lambda
  • aws-polly
  • aws-transcribe
  • catboost
  • cloudfront
  • cloudwatch
  • dynamodb
  • fastapi
  • framer-motion
  • kms
  • lightgbm
  • python
  • react
  • s3
  • scikit-learn
  • secrets-manager
  • shap
  • tailwind
  • typescript
  • vite
  • waf
  • xgboost
Share this project:

Updates