Agentic Emergency Guidance & Inter-agency Synthesis
Checking...
⚠ Railway platform degraded service. Status page →
Intake
Orchestrate
RAG
Bridge
Overseer
Synthesis
Routing Brief
Submit an incident report to generate a routing brief.
Agentic Architecture
1
Intake Agent
Parse & Classify
Reads the raw incident report and extracts structured intent — crisis type, location, state, and whether the input is complete enough to act on.
→ crisis_type · state · is_complete · raw_input
2
Orchestrator Agent
Route & Resolve
Routes the intent to the correct specialist cluster (Ohio / Oklahoma), assigns the regulatory agency by name, constructs the semantic RAG query, and loads the citation chain.
→ cluster · query · bbox · agency_routing · citation_chain
3
RAG Knowledge Agent
Retrieve
Performs semantic search over the policy document corpus (ChromaDB, local). Returns the most relevant context passages with a confidence score for the Overseer to validate.
→ context · citation · confidence
4
Data Bridge Agent
Live Data Fetch
Queries four live APIs in sequence: USGS real-time seismic feed, CDC Social Vulnerability Index by census tract, HHS emPOWER electricity-dependent beneficiaries, and EPA TRI hazmat facilities. Promotes EPA to Tier 1 if hazmat is detected.
→ usgs_live · svi_lookup · empower_data · epa_tri · agency_routing
5
IBM
Overseer Agent
Govern & Score
Runs three governance hooks at fixed pipeline positions. Scores each beam candidate by citation alignment — not token probability. Can trigger retry or force HONEST FALLBACK.
→ passed · citation_score · audit_log · retry_budget · output_state
6
IBM
Synthesis Agent
Generate & Select
Generates four beam candidates at varying temperatures. The Overseer selects the highest citation-aligned candidate — not the most probable token sequence.
→ routing brief · CONFIRMED DELIVERY · RETRY-CORRECTED · HONEST FALLBACK
IBM
Cards marked IBM use IBM Granite Guardian for input moderation and IBM Granite for text generation via IBM watsonx.ai.
Overseer Governance — Three Hooks
Hook 1
Input Audit
Fires after Intake. Flags unsafe, adversarial, or structurally incomplete input. Routes to HONEST FALLBACK immediately if failed — no downstream processing.
Hook 2
Retrieval Audit
Fires after RAG retrieval. Validates that retrieved context meets minimum confidence threshold. Triggers re-retrieval up to MAX_RETRIES before fallback.
Hook 3
Pre-Delivery Check
Fires after beam generation. Scores all candidates by semantic alignment with retrieved context. Selects the highest-scoring beam. Retries generation if no candidate passes.
Hybrid Archetypes: Knowledge Synthesizer Cognitive Load Reducer Human–AI Boundary Manager Decision Gatekeeper Invisible Coordinator
Governance & Trust
ImplementedExplainability

Every brief cites its sources. The Overseer scores each candidate by citation alignment before selecting one. The audit log records every governance hook with pass/fail and reason. Verification links point to the live USGS event and Census tract used to generate the output.

PartialTransparency

The dashboard surfaces pipeline status, citation alignment, retrieval confidence, and the full audit log on every run. This system is deployed in a professional EOC context where AI-assisted decision support is disclosed by policy. Model capability documentation follows IBM watsonx.ai platform standards.

ImplementedFairness

CDC Social Vulnerability Index (SVI) is the primary vulnerability lens — designed to surface disproportionate burden. HHS emPOWER adds electricity-dependent Medicare beneficiaries as a second equity dimension. The Ohio/Oklahoma geography was selected because induced seismicity concentrates in high-SVI tracts (Blackman, 2025).

ImplementedRobustness

Granite Guardian moderates user input. The Overseer runs three hooks — input audit, retrieval audit, pre-delivery check — with a retry budget of 2. HONEST FALLBACK is a first-class output state; the system never fabricates confidence it does not have. Beam search selects by citation alignment, not token probability.

PartialPrivacy

No user data is stored beyond the per-request audit log. All vector retrieval is local (ChromaDB on-server). Incident reports contain location and magnitude descriptions — no PII. Gap: the full prompt including retrieved context is transmitted to IBM watsonx for inference. For production EOC deployment, PII scrubbing on intake and a data processing agreement with IBM would be required.