PraxiCom

Archived portfolio project - built in spring 2025 for an AI in Healthcare course and the University of Texas at Austin School of Nursing's NURSING-AI Challenge. This repository is preserved as a proof of concept and is not under active development.

PraxiCom (formerly CareSpeak) is an affect-aware conversational simulation platform designed for APRN training. The system integrates real-time speech processing, structured persona modeling, and dynamic response generation to simulate emotionally nuanced patient interactions. It emphasizes communication complexity — including vocal affect, conversational resistance, and tone–content discrepancies — rather than diagnostic decision-making alone. The project was developed in collaboration with nursing faculty as part of the UT Austin NURSING-AI Challenge.

Student interface from the spring 2025 prototype.

Why I built it

Sensitive conversations - discussing end-of-life wishes, counseling an adolescent, or communicating a chronic diagnosis - require practice as much as clinical knowledge. Standardized-patient encounters are effective, but they are expensive, difficult to schedule, and hard to repeat.

The NURSING-AI Challenge invited graduate students and faculty advisors to build AI responses to real problems submitted by nursing faculty. PraxiCom was my response to Challenge 17, developed with faculty guidance from Dr. Cara Young. The prototype explored whether a faculty-controlled, synthetic AI patient could provide a scalable practice environment without using real patient data.

What the prototype demonstrates

• Real-time speech interaction through Hume AI's Empathic Voice Interface (EVI)
• Four faculty-oriented, YAML-authored patient scenarios with clinical context, emotional state, learning goals, and conversation guidance
• A student workspace for persona selection, session control, and faculty chat
• A faculty dashboard for joining by session ID, monitoring the transcript, quoting messages, and sending feedback
• A 10-criterion evaluation rubric with qualitative notes
• Local JSON export of the transcript, chat history, scores, and notes
• Session-end persistence in the faculty view so an evaluation can be completed after the student disconnects

Only the Advanced Directives scenario is configured for live voice in the archived prototype. The remaining personas document how the scenario library was intended to expand.

Architecture

flowchart LR
    Student["Student interface"] <-->|"audio, transcript, chat"| Relay["Node.js + Socket.IO relay"]
    Faculty["Faculty dashboard"] <-->|"monitoring, chat, evaluation"| Relay
    Relay <-->|"streaming voice session"| Hume["Hume EVI"]
    Personas["YAML patient personas"] --> Student
    Student -->|"selected persona + system prompt"| Relay

Loading

The browser streams microphone chunks to the relay server. The server owns session membership and forwards student audio to Hume EVI. Transcriptions, AI responses, synthesized audio, and faculty chat are broadcast through a session-specific Socket.IO room. Session data remains in memory for the life of the server; faculty exports are generated locally in the browser.

See docs/system-design.md for the event contract and module map.

Project structure

src/
  components/        Student/faculty interface components
  contexts/          Socket provider and context contract
  data/              Evaluation rubric
  hooks/             Socket and audio playback hooks
  pages/             Student and faculty routes
  services/          Persona loading and validation
  types/             Client domain types
server/src/
  services/          In-memory sessions and Hume EVI adapter
  socket/            Socket.IO event lifecycle
  types/             Server domain types
public/personas/     Synthetic patient definitions and portraits
docs/                Original challenge submission and design notes

Run locally

This is a legacy demonstration that depends on a Hume API account and an EVI configuration compatible with the archived SDK integration.

Prerequisites

• Node.js 18 or newer
• npm
• Hume API and secret keys

Install and configure

git clone https://github.com/rjleggieri/ut-nursing-ai.git
cd ut-nursing-ai
npm install
npm --prefix server install

Copy the example environment files and add your own credentials:

cp .env.example .env
cp server/.env.example server/.env

server/.env:

PORT=3001
CLIENT_URL=http://localhost:5173
HUME_API_KEY=your_hume_api_key
HUME_SECRET_KEY=your_hume_secret_key

Start the backend and frontend in separate terminals:

npm --prefix server run dev
npm run dev

Open http://localhost:5173. Start a student session, copy its eight-character ID, and use it from the Faculty Dashboard in a second browser tab.

Quality checks

npm run lint
npm run build
npm --prefix server test

Original submission

• Project report
• Challenge presentation
• Showcase handout

The original submission uses the project's earlier working title, CareSpeak. It describes the April 2025 prototype, its adaptation of GuideLLM concepts (goal navigation, context management, and empathetic engagement), and the ethical considerations around synthetic patients, student records, third-party AI services, HIPAA, and FERPA.

Scope and limitations

PraxiCom was an academic proof of concept, not a clinical product or validated educational intervention. It has no authentication, durable database, learning-management-system integration, or formal privacy/compliance controls. All included patients are synthetic. Do not use it with protected health information, real patient data, or student education records.

Modern multimodal AI platforms now provide much of this prototype's voice and orchestration capability directly. That makes the original product direction obsolete, but the project remains a useful record of the design problem, the 2025 implementation choices, and an early attempt to make generative patient simulation observable and faculty-directed.