Skip to content

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

🕒 Published (v1): 2026-03-25 22:51 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SentinelAI is a three-agent pipeline that converts heterogeneous emergency communications into NENA-compliant EIDO-JSON, correlates related reports across sources into composite incident views, and spatially enriches informal location descriptions. It operationalizes the newly adopted NENA-STA-024.1.1-2025 EIDO standard as an event-sourced, incrementally updated data model rather than a static exchange format. Enterprise interoperability is demonstrated via bidirectional FME integration.

Problem

Emergency response data arrives from multiple agencies (dispatch, fire, EMS, news, NWS) in heterogeneous formats with agency-specific codes, informal location references, and split temporal fields. Existing monolithic ETL pipelines cannot simultaneously handle language interpretation, temporal correlation, spatial enrichment, and semantic normalization at scale, nor do they align with the emerging NG9-1-1 EIDO standard for cross-agency interoperability.

Method

SentinelAI decomposes incident data integration into three asynchronous specialized agents:

  • EIDO Agent: Uses an LLM (model-agnostic, treated as a replaceable component) to extract entities from unstructured text, map agency codes to IANA registry values, synthesize timestamps into ISO 8601, and instantiate configurable EIDO-JSON templates per incident category.
  • IDX Agent: Correlates incoming EIDOs to existing incident contexts using a weighted tri-modal similarity score: $\(\Sigma(E_{new}, I) = w_t \cdot \phi_t(\Delta t) + w_g \cdot \phi_g(\Delta g) + w_s \cdot \phi_s(D_E, D_I)\)$ where \(\phi_t\) is temporal proximity, \(\phi_g\) is geospatial distance, and \(\phi_s\) is semantic similarity via sentence embeddings. A new EIDO is linked to incident \(I^* = \arg\max_I \Sigma(E_{new}, I)\) if \(\Sigma \geq \tau\); otherwise it initiates a new incident context. Composite views are computed dynamically: \(U_C(I) = \bigcup_{i=1}^{n} U(E_i)\) for resource sets; narrative fields are ordered by timestamp; state fields take the most recent value.
  • Geocoding Agent: Resolves informal location references by cascading through a curated local landmark index, then public mapping APIs (Google Maps, OSM), disambiguated by incident type and jurisdiction context.

FME integration uses a custom EIDOReader (EIDO-JSON → FME features) and EIDOWriter (structured FME data → NENA-compliant EIDO-JSON) for bidirectional enterprise data exchange.

Key Contributions

  • Reference architecture of three asynchronous specialized agents aligned with NENA-STA-024.1.1-2025.
  • Event-sourced incident model: EIDO objects are linked and never collapsed; composite views are derived on demand, preserving full audit trail.
  • Configurable EIDO templates maintained outside application code, enabling domain experts to adjust schema requirements per incident type without code changes.
  • Bidirectional FME integration (EIDOReader + EIDOWriter) demonstrating drop-in interoperability with existing geospatial/dispatch infrastructure.

Results

No quantitative benchmark results are reported. The paper presents a single illustrative case study: two San Diego County flood incident sources (NWS text alert + news article, January 1, 2026) are independently converted to EIDO-JSON (\(\text{EIDO}_A\), \(\text{EIDO}_B\)) and correctly linked by the IDX Agent based on temporal co-occurrence, spatial overlap (San Diego River within Central San Diego County polygon), and shared semantic terms ("flood", "rain"). No precision/recall, latency, or throughput figures are provided.

Limitations

  • No quantitative evaluation; no baselines, no recall/precision on incident correlation, no throughput benchmarks.
  • The threshold \(\tau\) and weights \(w_t, w_g, w_s\) are operationally configured but not learned or validated empirically; their sensitivity is uncharacterized.
  • Geospatial proximity uses straight-line distance by default; network-based alternatives are noted as optional but undemonstrated.
  • The LLM used for the EIDO Agent is not specified; extraction accuracy on adversarial or noisy inputs is not assessed.
  • Scope is limited to data structuring and interoperability; downstream decision support (routing, prioritization) is explicitly out of scope.
  • Single small case study; no multi-agency, high-volume, or real-time stress testing reported.

Relevance to Harnesses / Meta-Harnesses

SentinelAI exemplifies a domain-specific multi-agent harness where a fixed pipeline of specialized agents (ingest → correlate → enrich) is orchestrated asynchronously around a shared, schema-governed intermediate representation (EIDO-JSON). The use of externalized, expert-editable templates to control LLM output structure—decoupling prompt/schema configuration from application logic—is a concrete pattern for harness configurability. The event-sourced, incrementally-updated incident model (new EIDOs appended, composites derived on demand) parallels designs where a harness must handle streaming inputs and maintain a growing shared state without destructive merges. For researchers tracking harness design, the IDX Agent's weighted multi-signal correlation with a configurable threshold \(\tau\) demonstrates a reusable pattern for routing new artifacts to existing context buckets within a pipeline.