EMBERLINE 3.0 — From Chaos to Coordinated Action

Inspiration

When disaster strikes, help often arrives as chaos: a desperate text asking for insulin, a shelter running out of water, a family needing evacuation, and volunteers ready to help but unsure where they are needed most.

The problem isn't simply a lack of resources. It's the gap between receiving a request and coordinating the right response.

We built EMBERLINE to explore how artificial intelligence and mathematical optimization could help close that gap.

What it does

EMBERLINE 3.0 is an AI-powered disaster response coordination prototype that transforms unstructured requests into explainable, resource-aware relief plans.

Its end-to-end workflow includes:

• Multilingual AI triage: An NLP classifier processes requests in English and Greek, categorizing medical, water, food, transportation, shelter, and other needs.
• Information extraction: The system identifies locations, quantities, urgency, and cases requiring human review.
• Duplicate detection: Text similarity analysis highlights potentially repeated requests.
• Mathematical optimization: A 0–1 Mixed-Integer Linear Programming model matches needs with eligible volunteers while respecting their skills, coverage areas, availability, and remaining supplies.
• Human-in-the-loop coordination: Every proposed assignment requires explicit human approval before resources are reserved.
• Safety escalation: Potentially life-threatening situations are flagged and excluded from automated assignments.
• Scenario Lab: Users can simulate disruptions, including volunteer unavailability and supply shortages, then compare how the optimized response plan changes.
• Decision transparency: The dashboard explains assignment recommendations and records operator decisions in an audit trail.

The goal is not to replace emergency responders. It is to demonstrate how AI can support more transparent and coordinated human decisions.

How we built it

EMBERLINE combines natural language processing, operations research, and an interactive web dashboard.

Python and FastAPI power the backend. Machine learning techniques classify incoming messages, while similarity analysis detects potential duplicates. SciPy's HiGHS solver handles constrained assignment optimization.

The interface allows users to import synthetic requests, inspect AI predictions, generate optimized matches, approve proposed assignments, explore alternative scenarios, and review the resulting decisions.

We designed the optimization system to respect resource constraints, rather than simply matching each request to the first available volunteer.

Challenges we faced

One of the biggest challenges was making intelligent recommendations without sacrificing accountability.

A greedy assignment may look reasonable for one request but prevent a better overall allocation. We addressed this by introducing a mathematical optimization model with explicit constraints.

Another challenge was uncertainty. Humanitarian messages can be ambiguous, multilingual, incomplete, or extremely urgent. EMBERLINE therefore includes review mechanisms, explainable recommendations, and safety holds rather than treating every AI prediction as reliable.

Finally, we wanted the Scenario Lab to support experimentation without changing the underlying operational records.

Accomplishments we're proud of

• Built an end-to-end AI-assisted coordination workflow.
• Integrated exact optimization with interpretable assignment recommendations.
• Developed a Scenario Lab for comparing alternative resource conditions.
• Added human approval, safety escalation, and decision auditing.
• Passed 143 automated tests in the development environment.
• Verified the optimizer against exhaustive search on 80 randomized small-scale test cases.

These results validate specific aspects of the prototype under synthetic conditions, not effectiveness during real emergencies.

What we learned

Automation is most valuable when it increases human understanding and preserves accountability.

We learned how to connect probabilistic machine learning with deterministic optimization, how to make resource constraints visible, and how to design human intervention into the system rather than treating it as an afterthought.

What's next

Our next steps would be to evaluate the classifier on independently collected, appropriately consented data, expand language coverage, introduce verified geographic and travel-time information, and collaborate with qualified emergency-response organizations to assess practical requirements.

Before any real-world use, EMBERLINE would require extensive safety validation, security review, privacy protections, and operational testing.

EMBERLINE — From chaos to coordinated action.

Built With

• css
• fastapi
• highs
• html
• javascript
• machine-learning
• mixed-integer
• pytest
• python
• scikit-learn
• scipy
• sqlite