🤖 AI

SentinelAI: A Multi-Agent Framework for Structuring and Linking NG9-1-1 Emergency Incident Data

This paper introduces SentinelAI, a scalable multi-agent framework that transforms raw emergency communications into standardized, machine-readable NENA-compliant data to enable real-time integration, composite incident construction, and cross-source reasoning for Next Generation 9-1-1 systems.

Original authors: Kliment Ho, Ilya Zaslavsky

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kliment Ho, Ilya Zaslavsky

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the conductor of a massive, chaotic orchestra where every musician is playing a different song, in a different language, and on a different instrument. Some are shouting, some are whispering, and some are just humming. This is exactly what happens during a major emergency.

Police, fire departments, news reporters, weather services, and 9-1-1 call-takers all generate information about the same disaster, but they do it in their own unique ways. One might say "Code 3," another says "Urgent," and a third just writes a long paragraph about a flooded street.

SentinelAI is like a super-smart, multi-talented translator and organizer that steps in to make sense of this chaos. It doesn't just listen; it builds a single, clear, living story of the emergency that everyone can understand.

Here is how it works, broken down into simple parts:

1. The Problem: The "Tower of Babel" of Emergencies

In the old days, emergency systems were like walkie-talkies that only handled voice calls. Today, with Next Generation 9-1-1 (NG9-1-1), we have text, photos, videos, and sensor data pouring in from everywhere.

The problem is that this data is messy.

The Police might use a code like "211A."
The News might say "Heavy rain caused a flood."
The Weather Service might issue a "Flood Warning."

If a computer tries to read all these different formats at once, it gets confused. It doesn't know if "211A" and "Flood Warning" are talking about the same event.

2. The Solution: The SentinelAI Team

Instead of building one giant, clunky machine to do everything, the authors built SentinelAI, which is like a team of three specialized experts (called "Agents") working together. Think of them as a production crew for a live news broadcast.

🕵️ The EIDO Agent: The "Translator"

What it does: This agent takes the raw, messy input (like a frantic 9-1-1 call note or a weather text) and translates it into a standard, machine-readable format called EIDO-JSON.
The Analogy: Imagine a translator at the UN. If a delegate speaks in "Police Code," the EIDO Agent translates it into "Standard English" that every other system understands. It turns "Code 3" into "High Priority" and turns a messy date like "Yesterday at 3 PM" into a perfect computer timestamp. It ensures everyone is speaking the same language.

🔗 The IDX Agent: The "Detective"

What it does: This agent looks at all the translated reports and asks, "Are these about the same event?"
The Analogy: Imagine a detective solving a case. They have a report from a witness saying "I saw a fire at the park," and another report saying "Fire trucks are at the park." The IDX Agent connects the dots. It uses three clues to link them:
1. Time: Did they happen at the same time?
2. Place: Are they in the same neighborhood?
3. Story: Do the words match (e.g., both mention "fire")?
If the clues match, the agent links the reports together. Instead of having two separate files, it creates one Composite Incident—a single, evolving story that grows as new information arrives.

🗺️ The Geocoding Agent: The "Map Maker"

What it does: People often describe locations poorly. They might say, "The big red barn near the old oak tree" instead of giving a street address.
The Analogy: This agent is like a local guide who knows the neighborhood better than Google Maps. It takes vague descriptions like "near the stadium" and figures out the exact coordinates. It adds a precise pin to the map so emergency crews know exactly where to go, while still keeping the original description for context.

3. The Magic: A Living Story, Not a Static File

In traditional systems, once an incident is recorded, it's often stuck in a single file. If new info comes in, you have to overwrite the old file, and you might lose the history.

SentinelAI treats an emergency like a living document or a thread in a group chat.

Step 1: The Weather Service sends a warning. SentinelAI creates a "Story Start."
Step 2: A news report comes in 10 minutes later. SentinelAI doesn't delete the first story; it adds the new details to the thread.
Step 3: A police report arrives. It gets added to the thread too.

At any moment, you can look at the "Composite View" to see the full picture, or you can look at the "Thread" to see exactly how the story unfolded minute-by-minute. This is crucial because in emergencies, the situation changes fast.

4. The "Plug-and-Play" Feature (FME Integration)

The paper also shows that SentinelAI can plug into the software that cities and agencies already use (like FME).

The Analogy: Think of SentinelAI as a universal power adapter. You don't need to replace all your expensive appliances (the existing government systems); you just plug SentinelAI in, and it lets them talk to each other perfectly.

Why Does This Matter?

Before SentinelAI, emergency responders might be looking at three different screens with three different stories about the same flood. They might send a boat to the wrong place because the addresses didn't match.

With SentinelAI:

Clarity: All data is standardized.
Connection: Related reports are automatically linked.
Speed: New information is integrated instantly.

It turns a chaotic storm of data into a clear, unified map, helping humans and computers work together to save lives faster.

1. Problem Statement

Emergency response systems generate vast amounts of data from diverse sources (police, fire, EMS, news, government) in varying formats (structured fields, free-text, audio, images). Integrating this data into a unified, machine-readable view is challenging due to:

Heterogeneity: Data arrives incrementally from different agencies with conflicting terminologies, coding schemes (e.g., "Code 3" vs. "P1"), and formatting standards.
Lack of Standardization: Existing monolithic pipelines struggle to simultaneously handle language interpretation, temporal correlation, spatial ambiguity, and semantic normalization.
NG9-1-1 Complexity: The transition to Next Generation 9-1-1 (NG9-1-1) increases data volume and diversity (text, video, sensors), requiring a system that can treat incidents as evolving events rather than static records.
The Gap: There is a need for a framework that can transform raw, heterogeneous communications into the NENA Emergency Incident Data Object (EIDO) standard while dynamically linking related reports from multiple sources into a composite incident view.

2. Methodology: The SentinelAI Framework

SentinelAI addresses these challenges using a multi-agent architecture that decomposes the complex data structuring task into discrete, specialized sub-tasks. The system operates asynchronously, allowing agents to function independently while contributing to a unified incident representation.

The framework relies on three core agents and one integration layer:

A. The EIDO Agent (Structuring & Interpretation)

Function: Acts as the primary ingestion point, transforming heterogeneous inputs (call-taker notes, radio transcripts, dispatch summaries) into standardized EIDO-JSON documents compliant with the NENA-STA-024.1.1-2025 standard.
Mechanism:
- Utilizes Large Language Models (LLMs) to interpret unstructured text and extract entities (people, locations, organizations).
- Normalization: Maps proprietary agency codes (e.g., "211A", "Code 3") to standardized IANA registry values.
- Temporal Synthesis: Combines disparate date/time fields into ISO 8601 timestamps.
- Template-Based: Uses configurable templates to define required/optional fields for specific incident types, allowing subject-matter experts to adjust logic without code changes.

B. The IDX Agent (Correlation & Synthesis)

Function: Links new EIDO documents to existing incident contexts or initiates new ones, creating a "composite incident" from multiple partial reports.
Mechanism:
- Treats an incident $I$ as a set of linked EIDO objects $\{E_1, E_2, ..., E_n\}$ .
- Scoring Algorithm: When a new EIDO ( $E_{new}$ ) arrives, it calculates a similarity score $\Sigma$ against existing incidents using a weighted function:
  $\Sigma = w_t \cdot \phi_t(\Delta t) + w_g \cdot \phi_g(\Delta g) + w_s \cdot \phi_s(D_E, D_I)$
  Where $\phi_t$ is temporal proximity, $\phi_g$ is spatial distance, and $\phi_s$ is semantic similarity (via vector embeddings).
- Thresholding: Links the new EIDO to the incident with the highest score if it exceeds a configurable threshold $\tau$ . If no match is found, a new incident context is created.
- Composite View: Dynamically constructs a unified view of the incident by aggregating data from all linked EIDOs (e.g., union of all units, most recent status, chronological narrative) without collapsing the underlying data into a single mutable record.

C. The Geocoding Agent (Spatial Enrichment)

Function: Resolves informal location descriptions (landmarks, colloquial names, "near the market") into structured spatial data.
Mechanism:
- Combines local knowledge bases (curated landmarks), external mapping APIs (Google Maps, OpenStreetMap), and incident context (jurisdiction, incident type) to disambiguate locations.
- Outputs multiple location forms in the EIDO: coordinates, matched landmark names, descriptive text, and civic addresses, along with a confidence score.

D. Enterprise Integration (FME)

Function: Demonstrates interoperability with existing enterprise systems using Safe Software's FME (Feature Manipulation Engine).
Components:
- EIDOReader: Converts EIDO-JSON into FME features for spatial analysis and mapping.
- EIDOWriter: Converts structured legacy data (spreadsheets, DBs) into valid EIDO-JSON for ingestion into SentinelAI.

3. Key Contributions

Three-Agent Reference Architecture: A concrete implementation of a scalable, asynchronous multi-agent system aligned with the NENA EIDO standard, separating concerns into ingestion (EIDO), correlation (IDX), and enrichment (Geocoding).
Operational Linked-Incident Data Model: A methodology treating EIDO as an evolving, event-sourced representation. Instead of overwriting a single record, the system links incremental updates to a common context, preserving the full history and provenance of the incident.
Enterprise Interoperability: A demonstrated proof-of-concept showing how EIDO-JSON can flow bidirectionally between the SentinelAI framework and existing geospatial/emergency management workflows via FME.

4. Results (Case Study)

The paper validates the framework using a realistic scenario involving heavy rainfall in San Diego County (Jan 1, 2026):

Input: Two distinct sources—a National Weather Service (NWS) text warning and a San Diego Union-Tribune news article.
Process:
- The EIDO Agent converted the NWS warning and the news article into separate EIDO-JSON documents, extracting event types ("Flood Warning") and locations ("Central San Diego County" vs. "San Diego River").
- The Geocoding Agent resolved "Central San Diego County" to a polygon and "San Diego River" to a specific linear feature.
- The IDX Agent calculated high similarity scores based on temporal proximity (same day), spatial overlap (river within the county polygon), and semantic overlap (keywords: "flood," "rain").
Outcome: The system successfully linked the two disparate reports into a single composite incident timeline. The resulting view combined the official warning status with specific impact details (roadway flooding, power outages), providing a more complete operational picture than either source alone.

5. Significance

Standardization: Provides a practical pathway for agencies to adopt the NENA EIDO standard, moving away from siloed, proprietary data formats.
Scalability & Flexibility: The multi-agent design allows for the independent scaling of specific functions (e.g., upgrading the LLM for the EIDO Agent without disrupting the IDX logic) and handles the "messy" reality of emergency data (informal language, incremental updates).
Traceability: By preserving individual EIDO objects rather than merging them into a single static record, the system maintains a complete audit trail of how an incident was understood and updated over time.
Interoperability: The FME integration proves that this advanced AI framework can coexist with legacy government infrastructure, facilitating a smoother transition to NG9-1-1 ecosystems.

In summary, SentinelAI offers a robust, standards-aligned solution for transforming fragmented emergency communications into structured, linked, and actionable intelligence, directly addressing the data engineering challenges inherent in modern Next Generation 9-1-1 systems.