What challenges did RapidSOS face before AI implementation?

Emergency dispatchers were overwhelmed by high volumes of non-emergency 911 calls. The system needed real-time automation that could safely handle routine calls while escalating possible emergencies immediately.

What was the impact of the AI solution?

The article reports that the system handles 60-70% of non-emergency calls without human intervention. It is optimized for sub-100ms performance in the core loop. The dual-model verification approach improves safety by filtering false positives and escalating true emergencies quickly.

Source

GCPProductionEvidence: Medium50/100

RapidSOS built HARMONY voice agent for non-emergency 911 calls using Gemini Live API + ADK

Use case typeVoice automationUpdated Jun 3, 2026

RapidSOS, a public safety AI company, built HARMONY AI as an intelligence layer for 911 centers to automate non-emergency calls and support dispatchers. The voice agent dynamically guides callers through agency-specific protocols, verifies critical location data, packages incident details for downstream field processing, and escalates to a human dispatcher when emergency risk is detected. The system uses Gemini Live API on Gemini Enterprise Agent Platform, Google Agent Development Kit (ADK), Gemini 2.5 Flash, and Chirp 3 in an agentic architecture with parallel supervision and safety verification.

Organization: RapidSOS
Industry: Public Sector
Location: United States
Published: June 2026

Reported outcomes

100 ms

core loop latencyTime & speed

60-70%non-emergency call completion without human intervention

Strategic outcomes

Risk & complianceSafer automated escalation for possible emergenciesCost efficiencyFreed dispatchers for life-saving emergencies

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

Non-emergency call completion without human intervention: 60-70% increase

Google Developer BlogJun 3, 2026Blog postExplicit claimMedium evidence strength

Successfully handles 60-70% of non-emergency calls without human intervention.

Normalized claim

Core loop latency: 100 ms decrease

Google Developer BlogJun 3, 2026Blog postExplicit claimMedium evidence strength

Sub-100ms performance.

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: RapidSOS
Provider: GCP
Maturity: Production
Linked source: Google Developer Blog

RapidSOS deployed a purpose-built voice agent for non-emergency calls

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

1Voice automation
2Safety monitoring

Emergency dispatchers were overwhelmed by high volumes of non-emergency 911 calls.
The system needed real-time automation that could safely handle routine calls while escalating possible emergencies immediately.

RapidSOS deployed a purpose-built voice agent for non-emergency calls.
The implementation uses Gemini Live API for low-latency audio-to-audio conversation, Google Agent Development Kit (ADK) for the conversational loop, Chirp 3 for parallel transcription, and a secondary Gemini supervision model to monitor audio and transcripts in parallel.
A dual-model voter/supervisor pattern cross-references critical keywords and caller sentiment, triggers self-healing conversation behavior when calls stall, and routes risky calls to a human dispatcher.

Technologies

Gemini Live API Google Agent Development Kit (ADK)Gemini Enterprise Agent Platform Gemini 2.5 Flash Chirp 3

The article reports that the system handles 60-70% of non-emergency calls without human intervention.
It is optimized for sub-100ms performance in the core loop.
The dual-model verification approach improves safety by filtering false positives and escalating true emergencies quickly.

Architecture

RapidSOS built HARMONY AI as an intelligence layer for 911 centers. The architecture uses Gemini Live API on Gemini Enterprise Agent Platform as the primary audio-to-audio interface, Google Agent Development Kit (ADK) for the conversational loop, Chirp 3 as a parallel transcriber, and a secondary Gemini 2.5 Flash supervision model that monitors audio and transcripts in parallel to verify safety and escalate when needed. The system also includes a dual-model voter pattern, keyword and sentiment checks, and self-healing conversation handling.

Sources & evidence1

Evidence: Medium50/100Evidence strength