An AI-powered wilderness rescue simulator with human-in-the-loop validation.
https://summit-sim.bhamm-lab.com/
Summit-Sim generates curriculum-informed, interactive backcountry emergencies for Wilderness First Responder (WFR) training. Instructors review and approve AI-generated scenarios before students engage with them, ensuring medical accuracy and pedagogical value.
Built for the Weber State AI Hackathon 🐾
Acknowledgments: Thanks to Keenan Grady at the Ogden Avalanche Center for domain expertise in wilderness safety and WFR curriculum.
Wilderness First Responder (WFR) training relies heavily on static paper scenarios or expensive live-action roleplay. Students rarely get enough dynamic, unpredictable repetitions to truly test their decision-making under pressure. When emergencies happen in the backcountry, textbook memorization isn't enough—responders need dynamic critical thinking developed through varied practice.
Summit-Sim provides infinite, medically accurate WFR scenarios through a dynamic AI game loop. Instead of multiple-choice questions, students use natural language to evaluate scenes, check vitals, and apply treatments. The system evaluates every action against a hidden medical "truth," evolving the scenario exactly as a real patient's condition would change.
- Infinite Scenario Generation: Create highly specific emergencies based on environment, group size, difficulty, and complexity. Each scenario includes a unique AI-generated atmospheric image.
- Human-In-The-Loop (HITL) Validation: Instructors review scenarios before publication, with all feedback logged to MLflow for continuous improvement.
- Natural Language Interaction: Students type free-text actions (no multiple choice). The AI evaluates against WFR protocols and reveals information progressively.
- Cumulative PAS Scoring: Tracks student progress (0-100%) across 5 Patient Assessment System milestones: Scene Size-up, Primary Assessment, Secondary Assessment, Treatment, and Evacuation Plan.
- Intelligent Debriefing: Post-simulation analysis with clinical reasoning assessment, key mistakes identification, and personalized recommendations.
- Shareable Scenarios: Students join via unique URLs—multiple students can work on the same scenario in different sessions.
- Solo Practice Mode: Students can also generate their own scenarios for independent practice. The scenario auto-generates and immediately starts the simulation without instructor review. Students never see hidden information (learning objectives, hidden truth, or hidden state)—they must discover medical details through proper assessment, just like in the instructor-led flow.
Summit-Sim is built with a sophisticated AI stack focused on medical safety, strict output schemas, and comprehensive observability.
| Component | Technology | Purpose |
|---|---|---|
| Orchestration | LangGraph | Complex state flows, HITL interrupts via interrupt(), checkpointing for persistence |
| Agent Framework | PydanticAI | Strict medical safety via enforced Pydantic schemas for all LLM I/O |
| UI Framework | Chainlit | Async, reactive Python UI for conversational flows |
| Observability | MLflow | LLM span tracing, human feedback tracking, prompt versioning, GEPA optimization |
| State Storage | Redis (local) / DragonflyDB (prod) | LangGraph checkpoint persistence and scenario storage |
| Image Generation | OpenRouter (Gemini Flash Image) | Unique atmospheric scenario images (16:9, non-blocking) |
The application runs two interconnected LangGraph workflows:
- The Authoring Graph: Generates scenarios from environmental parameters, creates atmospheric images, and presents them for instructor review via HITL interrupt before publication. When students generate their own scenarios, this graph auto-approves and bypasses the review interrupt, jumping directly to simulation.
- The Simulation Graph: A continuous game loop where students type actions, the AI evaluates against hidden medical truth, and the patient state evolves dynamically. Completes at 80% PAS milestone completion or max turns. All students (whether joining via instructor link or generating solo) see only observable information—hidden medical data is never revealed without proper assessment.
Four specialized PydanticAI agents power the system:
| Agent | Purpose | Output Schema |
|---|---|---|
| Generator | Creates wilderness rescue scenarios from configuration | ScenarioDraft with visible/hidden separation |
| Image Generator | Produces unique scenario images | Base64-encoded 16:9 image |
| Action Responder | Evaluates student actions, reveals information, updates PAS score | ActionResponse with narrative and feedback |
| Debrief | Generates post-simulation performance analysis | DebriefReport with teaching points |
MLflow Judges (Implemented): Four specialized judges evaluate simulation quality:
- Structure Judge: Score range validation, narrative length, feedback tone
- Scoring Judge: Milestone justification, monotonic progress, acknowledgment
- Medical Judge: Treatment gate validation (was_correct accuracy)
- Continuity Judge: Progressive revelation, score monotonicity
Note: Judges are implemented but currently disabled in production due to MLflow bug #20782. They are used for offline GEPA (Genetic Pareto) optimization in Jupyter notebooks.
GEPA Optimization: Uses the judge framework with reflection models to iteratively improve agent prompts based on expert feedback. See notebooks/action-responder-prompt.ipynb.
WFR Curriculum references used to ensure medical accuracy:
- https://www.scribd.com/document/38484292/Wilderness-First-Responder-Course
- https://wildsafe.org/wp-content/uploads/2021/11/CWS-%E2%80%93-WFR-Partcipant-Workbook-2021.pdf
- https://www.nols.edu/category/wilderness-medicine/case-studies/
- https://www.nols.edu/category/wilderness-medicine/case-studies/page/2/
- https://www.letsgoexploring.com/public/wilderness-first-responder-infosheet-side-A-2022.pdf
- https://www.wildmedcenter.com/uploads/5/9/8/2/5982510/standard_wfr_syllabus.pdf
- https://sierrarescue.com/wp-content/uploads/2013/12/WILDERNESS-FIRST-RESPONDER-General-Info.pdf
- https://sierrarescue.com/coursepdf/WFR.pdf
- https://www.nols.edu/wp-content/uploads/2025/08/23732-Student-Logistics.pdf
- https://www.deepsprings.edu/next/wp-content/uploads/2024/07/Student-WFR-Training.pdf