Praxis

Inspiration

Medical education has a critical problem: physicians need thousands of hours of patient interactions to develop strong clinical reasoning skills, but standardized patient actors are expensive, limited, and don't scale. Meanwhile, advanced AI models like Kimi K2.5 have sophisticated reasoning capabilities that rival expert physicians.

We asked: What if we could create an accessible, AI-powered platform where medical students practice diagnostic thinking with an intelligent patient that responds naturally to their questions?

Praxis brings medical education into the 21st century by combining:

Advanced AI reasoning (K2.5) that understands medical nuance
Real medical literature (PubMed) for realistic, evidence-based cases
Natural voice interactions (ElevenLabs TTS + AssemblyAI) for immersive practice
Socratic method pedagogy that guides students toward diagnosis rather than telling answers

What It Does

Praxis is an intelligent medical education platform that trains physicians and medical students in clinical reasoning through immersive, AI-powered patient interviews.

Core Features

1. Interactive Patient Interviews

Real-time conversations with an AI patient that roleplays clinical scenarios
Powered by Kimi K2.5 for sophisticated, medically-accurate responses
Students interact via text or voice (with AssemblyAI transcription)
AI responds in natural voice via ElevenLabs TTS
Socratic method teaching guides students toward correct diagnosis

2. Literature-Based Case Generation

Automatically generate realistic clinical cases from peer-reviewed medical literature
Searches PubMed via NCBI API for real case reports
Synthesizes cases using K2 based on actual medical literature
Cases include differential diagnoses, red flags, and learning objectives
All cases cite actual PMIDs for literature grounding

3. Demo Cases for Immediate Practice

5 pre-built clinical scenarios requiring zero setup:
- Acute Myocardial Infarction
- Pneumonia with Consolidation
- Acute Appendicitis
- Acute Ischemic Stroke
- Diabetic Ketoacidosis
Ability to generate new clinical scenarios on the fly

4. Session Management & Progress Tracking

Track all past consultations with timestamps
View interaction history and AI feedback
Session evaluation with performance metrics
Both in-memory (demo) and database-persisted sessions

How We Built It

Tech Stack

Frontend: React 18 + TypeScript + Vite + Tailwind CSS + React Router + Radix UI

Backend: FastAPI (Python) + Supabase PostgreSQL + Pydantic validation

AI & APIs:

Kimi K2.5 (via Featherless) - Advanced reasoning for patient roleplay
ElevenLabs TTS - Natural voice generation
AssemblyAI - Audio transcription for voice input
NCBI PubMed API - Literature search for case generation

Architecture

Frontend (React/Vite)
    ↓
REST API (FastAPI)
    ├─ /api/sessions - Session management
    ├─ /api/reasoning - Clinical analysis
    ├─ /api/dedalus - Case generation
    └─ /api/consultations - Interview interactions
    ↓
Supabase Database (PostgreSQL)
    ├─ Cases
    ├─ Sessions
    ├─ Interactions
    └─ Users
    ↓
External APIs
    ├─ Kimi K2.5 (Featherless)
    ├─ ElevenLabs TTS
    ├─ AssemblyAI
    └─ NCBI PubMed

Implementation Highlights

Tool calling for structured PubMed searches within K2 reasoning
In-memory sessions for demo cases + database persistence for real cases
Lazy initialization pattern for external API services
Async/await throughout for responsive frontend
Real-time voice pipeline: Audio capture → AssemblyAI → K2 reasoning → ElevenLabs TTS → Playback

Challenges We Ran Into

1. Demo vs Database Session Handling

Problem: Demo cases have string IDs (like "case-1") but the backend enforced foreign key constraints requiring UUIDs in the database. This caused HTTP 422 errors when starting demo case sessions.

Solution:

Added UUID validation helper to distinguish demo (strings) from real cases (UUIDs)
Created in-memory session objects for demo cases
Skip database operations for demo sessions
Return generic greetings for non-database cases

2. K2 JSON Output Reliability

Problem: K2 was returning analysis/thinking text before JSON instead of pure JSON output.

Solution:

Simplified prompt with explicit "GENERATE JSON ONLY" instruction
Added robust JSON extraction with fallback parsing
Strip <think> tags from K2 responses
Implemented graceful error handling

3. PubMed API Returns XML, Not JSON

Problem: Code requested rettype=json but NCBI API returns XML by default, causing parse errors.

Solution:

Switched to XML parsing using xml.etree.ElementTree
Extract real article PMIDs from XML responses
Removed all fallback dummy data—cases now require actual peer-reviewed literature

4. Variable Shadowing Bug

Problem: A status variable was overwriting the imported status module from FastAPI, causing AttributeError.

Solution:

Renamed session status variable to session_status
Proper scope management prevents module shadowing

5. Case Model Validation

Problem: CaseCreate expected clinical_scenario as Dict but Dedalus generated string values.

Solution:

Updated model to clinical_scenario: str (the patient's opening speech)
Made differential_diagnoses and red_flags optional
Flexible schema supports both demo and generated cases

Accomplishments That We're Proud Of

Built a fully functional medical education platform with real-time AI interactions
Integrated K2.5 reasoning engine for sophisticated Socratic responses
Connected to actual medical literature (PubMed) for authentic case generation
Implemented complete voice pipeline (capture → transcription → reasoning → TTS → playback)
Fixed critical session management bugs - Demo and database cases now work seamlessly
Created 5 realistic demo cases ready for immediate use with zero setup
Built authentication system with Supabase for user accounts and progress tracking
Production-ready code - All changes tested and deployed

End-to-End Test Results

Demo cases: Session creation (HTTP 201), patient interaction (HTTP 200)
Audio generation: 87-119KB MP3 files generated and playable
Frontend: 1.48s build time, 1793 modules, zero TypeScript errors
API responses: <500ms average latency
No database errors or session management issues

What We Learned

AI in Medical Education Works - K2.5's reasoning is sophisticated enough to hold realistic medical conversations and guide students toward diagnostic thinking.
Literature-Grounded AI is Essential - Cases grounded in real PubMed articles are far more credible and effective than synthetic examples.
The Socratic Method Scales - We can automate the "question-guiding-student" pattern at scale, reducing reliance on expert educators.
Demo vs Real Data Strategy - Providing instantly-usable demo cases while supporting persistent database cases created the best user experience balance.
Voice Interactions Are Powerful - The combination of TTS (AI speaks) + transcription (student speaks) creates a more natural learning experience than text-only.
External APIs Require Defensive Design - Integrating with PubMed, ElevenLabs, and AssemblyAI taught us to expect different data formats and build robust fallbacks.
Tool Calling Unlocks Structured Reasoning - K2's ability to call PubMed search tools within reasoning loops enables dynamic case generation.

What's Next for Praxis

Short Term (Current Sprint)

[ ] Add case evaluation metrics and scoring system
[ ] Implement real-time progress dashboard
[ ] Add 10+ additional demo cases across specialties
[ ] Improve Dedalus JSON reliability with retries

Medium Term (Q2 2026)

[ ] ChromaDB integration for semantic case search
[ ] Case difficulty auto-adjustment based on student performance
[ ] Peer leaderboards for friendly competition
[ ] Video integration for physical exam practice
[ ] Mobile app version (iOS/Android)

Long Term (Q3+ 2026)

[ ] Spaced repetition scheduling for optimal retention
[ ] Personalized learning paths by medical specialty
[ ] Integration with medical school curricula
[ ] AI-powered performance analytics and recommendations
[ ] Board exam preparation mode with full-length mock exams
[ ] Multi-patient scenarios (simultaneous patient management)

Impact & Use Cases

For Medical Students: Practice clinical reasoning 24/7, get immediate feedback, build confidence before real patient encounters

For Medical Schools: Scalable training platform, track competencies, reduce cost of standardized patient actors, support remote/hybrid education

For Residents: Maintain skills, prepare for board exams, specialty-specific case practice

For CME: Continuing medical education that adapts to learning needs

Try It Out

Quick Start

# Frontend
npm install && npm run dev  # http://localhost:3002

# Backend
cd backend
pip install -r requirements.txt
python -m uvicorn app.main:app --reload  # http://localhost:8000