Crisis Forge

CrisisForge is an initiative to improve simulations training, particularly for better inter-organizational collaboration during crises, and to create tools for use during crises.

This project started as a Hackathon hosted by UBC's Emerging Media Lab with mentorship provided by members of the Justice Institute of British Columbia and Conversa Corps. With a focus on JIBC's "Praxis" tabletop simulation tool, participants made a wide range of projects ranging from UI/UX designs to functional prototypes:

---

1. Nonlinear Crisis Timeline Interface with AI-Assisted Injects

• Problem: Traditional crisis management timelines are too linear and hard to modify dynamically.
• Solution: A redesigned user interface for the PRACTICE system that supports branching timelines and AI-generated injects.
• Features:
  • Drag-and-drop inject reordering.
  • AI-generated visuals for injects.
  • Enhanced buttons, added “Cancel” and “Back to Dashboard.”
  • Categorization and sub-inject numbering system (e.g., 2.1, 2.2).
• Goal: Improve usability and realism of crisis simulations, especially for administrators and controllers.

---

2. AI-Supported Debriefing Tool for Decision Traceability

• Problem: Experts only see the final decision and rationale, missing insights from the full conversation.
• Solution: Use of LangChain and AI prompts to analyze chat logs and generate:
  • All possible decisions discussed.
  • Contributions of individual participants.
  • A percentage breakdown of how much each comment influenced each decision.
• Prototype Features:
  • Review reports with full chat history.
  • AI-generated potential decisions.
  • Linked rationale and contributors per decision.
• Future Potential: Applying this to tools like Slack or Zoom for structured insight capture.

---

3. Dynamic Crisis Simulation Engine with Recursive AI Prompts

• Problem: Need for responsive, iterative crisis simulations that evolve with participant input.
• Solution: A modular system that separates simulation into:
  • Baseline prompt (overall situation).
  • Rolling situational updates.
  • Ongoing user assessments and actions.
• Architecture:
  • Containerized microservices (for scalability).
  • Uses prompt chaining and AI evaluation to simulate crisis evolution.
  • Efficient token usage and data storage.
• Outcome: A scalable, Zork/AI Dungeon-style real-time crisis interaction system.

---

4. Real-Time Slide Deck with Live Data and GPT Integration

• Problem: Static slides lack adaptability in crisis simulations.
• Solution: A prototype for slide decks that dynamically pull real-time data and GPT outputs per slide.
• Use Case: Command centers or practice simulations with evolving real-world info (e.g., wildfire updates).
• Features:
  • Slides auto-refresh with live or pseudo-live data (e.g., from Wikipedia or news).
  • Future vision includes live inputs from websites or APIs to drive simulation updates.
• Challenges: Debugging inconsistent live data rendering.

---

5. After Action Report (AAR) Generator + Decision Dynamics Evaluator

• Problem: Generating quality AARs is time-consuming and subject to human oversight.
• Solution:
  • Use GPT to generate AARs based on FEMA templates.
  • Evaluate team decision-making dynamics from transcripts.
• Experiments Included:
  • Testing with fictional input (e.g., Star Trek, Council of Elrond) to simulate decisions.
  • Added a fictional disruptive participant to evaluate how the system handled poor input.
• Result: Dynamic scoring, feedback, and insights into group performance and individual contributions.

---

6. Sphere Handbook RAG (Retrieval-Augmented Generation) Assistant

• Problem: Humanitarian guidelines are complex and hard to access in real-time.
• Solution: LangChain-based tool that uses a vector database of the Sphere Handbook (~450 pages).
• Features:
  • AI answers questions like “How do we get water to isolated communities?”
  • Cites specific handbook segments and page references (with some formatting issues to fix).
• Goal: Reliable, non-hallucinated AI access to expert humanitarian protocols.

---

Additional Contributions and Observations:

• Early GPT-based scenario generation tool: Prompt-based assistant trying to generate JSON scenarios (MVP level).
• A few teams collaborated closely and learned tools like LangChain, vector stores, prompt engineering, and GPT-based API integration on the fly.
• Suggestions for future work:
  • Integration with RAG (Retrieval-Augmented Generation) pipelines for knowledge bases.
  • Incorporating Reinforcement Learning (e.g., RH models) for more accurate simulations.
  • Expanded containerization and microservice architecture for scale and deployment.

---

General Themes:

• Emphasis on nonlinear thinking, human-centered interface design, and modular AI pipelines.
• Strong orientation toward applicability in both simulations and real-world crisis scenarios.
• Tools developed were portable, flexible, and grounded in practical crisis management needs.

---

Presentations from the Hackathon may be seen here: https://www.youtube.com/watch?v=DTsG-YLgj1M