Solace — AI-native ED intake & triage

The description states that Solace is a two-sided system built around a shared backend on Amazon DynamoDB:

• Patient front door: A multilingual, voice-based intake process accessible via QR code at the ED entrance. Patients can speak or type in plain language, describe symptoms, and optionally upload images. The system provides an AI-generated triage priority spoken back to them.

• Clinician cockpit (Atlas): A single-screen terminal that displays live patient queues ranked by ESI (Emergency Severity Index), with clinical notes, scribes, copilot tools, EHR integration, and workflow automations.

The system is built using AWS services including Lambda, Bedrock, DynamoDB, CloudFront, and Vercel. It uses a stacked ensemble of machine learning models for triage, with SHAP and conformal prediction for explainability.

Evidence: The description states this is the product’s architecture and functionality.

The author positions Solace as a solution to inefficiencies in ED workflows:

• Reducing rework by eliminating redundant data entry.
• Addressing language barriers that lead to under-triaging.
• Eliminating clinicians' need to reconstruct patient stories from multiple sources.
• Providing a system that removes “dead time” from the front end of an ED encounter.

The product is described as not being a chatbot, but rather a system that integrates intake and triage into one shared source of truth. It emphasizes speed, multilingual support, and integration with EHRs and clinical workflows.

Evidence: The description states this is the intended positioning and evolution of the product.

The description identifies emergency departments (EDs) as the primary target customer. The system is designed for use in hospitals where EDs are bursty, unpredictable, and require fast, scalable systems.

It also implies a dual user base:

• Patients: ED visitors who interact via voice or text.
• Clinicians: ED staff using the Atlas cockpit to triage and manage patient encounters.

The author notes that the system is built for hospitals with limited English proficiency patients (roughly 1 in 5 visits), suggesting a focus on urban, diverse healthcare environments.

Evidence: The description states this is the intended target customer and ICP.

Not evidenced. The description does not mention pricing, licensing models, or any commercial structure beyond what was built for a hackathon.

The system is built on AWS with:

• Database: Amazon DynamoDB (multi-table, on-demand, CMK-encrypted)
• Frontend: Vercel (SPA and static deployment)
• API Layer: FastAPI on AWS Lambda
• AI Services: AWS Bedrock (Claude), Transcribe, Polly
• ML Stack: Stacked ensemble of LightGBM, XGBoost, CatBoost, MLP with SHAP and conformal prediction
• Security: AWS WAFv2, CloudTrail, IAM, SMART-on-FHIR authentication

The system is designed to be fast (single-digit-millisecond reads), scalable (no capacity planning), and HIPAA-compliant.

Evidence: The description states these are the technical components and design decisions.

Not evidenced. There is no mention of customers, revenue, usage metrics, or any real-world deployment beyond the hackathon submission.

Not evidenced. No information is provided about competitors or market positioning in relation to existing ED triage or intake systems.

• Unproven clinical validity: The system claims to use AI for triage but does not provide evidence of clinical validation or testing.
• No real-world deployment: The product is described as a hackathon submission, with no evidence of being used in hospitals.
• Self-reported safety and compliance: While it mentions HIPAA compliance and AWS BAA, there is no independent verification of these claims.
• Single-person team: The system is built by one person (Dhruv Jain), raising questions about scalability and operational capacity.

Inference: The lack of real-world testing or validation makes the product risky for commercial adoption.

1. Has this system been tested in a real ED setting?
2. What clinical outcomes have been observed from using this triage system?
3. How does the system handle edge cases or ambiguous patient inputs?
4. Are there any partnerships with hospitals or healthcare systems currently in place?
5. What is the plan for scaling beyond a single hospital or region?
6. How is the model retrained or updated to reflect new clinical guidelines?