1Waiting
Signal agent
Read messages and extract explicit facts with source IDs.
Awaiting input
Waiting to run
THREE-AGENT AI INCIDENT BRIEFING
Responder Incident Brief.
A three-agent Unified Communications AI concept: read scripted radio-style text, extract stated facts with source IDs, check proposed facts against the original messages, and draft a short update for a responder joining an incident. A person reviews the draft; the original communications remain available when AI fails.
6Scripted messages
9m 22sScenario timeline span
1Unclear messages
DemoProvider mode
Incoming incident event streamScenario loaded · ready to receive messages
Scenario queued · press Run the workflow
PROVIDER
Offline workflow ready · no key needed
0Event intakeTime · speaker · source ID
1SignalExplicit facts
2SafetyFact-level review
3BriefingAdvisory draft
4Responder handoffSources · human review
Incoming transmissions STEP 0
0 transmissionsScenario ready · press Run the workflow.
AGENT ORCHESTRATION · LIVE TRACE
Agent workflow STEPS 1-3
Three bounded agents hand off evidence. None can dispatch, route calls or approve its own output.
Offline simulation
Incoming incident transmissions0 received
Scenario queued · press Run the workflow.
Connected mode sends visible synthetic message text directly from this browser to the selected provider. The API key stays in page memory for this session and is cleared on reload. Connect saves it in memory; the workflow tests the first actual API call. Never use real incident data.
2Waiting
Safety agent
Check evidence and uncertainty. Withhold claims when context is unsafe.
Awaiting input
Waiting to run
3Waiting
Briefing agent
Draft a short catch-up from approved facts for human review.
Awaiting input
Waiting to run
Scenario ready. Press Run the workflow to receive the messages and run each stage.
FINAL RESPONDER HANDOFF
Advisory briefing ready for review
Review source messages with dispatch before any operational decision. This demo does not dispatch or route calls.
Advisory briefing STEP 4
Waiting for contextNo briefing yet. Press Run the workflow to watch messages arrive and agents build a sourced draft.
Trust and control
0 sources0 uncertainHuman verification required
The briefing is advisory. It never routes calls, dispatches units, or replaces the original message. When AI is unavailable, the feed remains visible.
CUSTOMER DESIGN / REFERENCE ARCHITECTURE
It advises. It does not act.
Built for what happens when it fails. Three agents read the radio traffic and write a short briefing. Each fact is checked against the original message. If a responder says he is not sure, the agent leaves it out. If the AI stops working, the messages are still there.
What this demo is: synthetic voice clips with transcripts written in advance. A real service would need permission to use agency audio, speech-to-text that has been tested and measured, policy controls, and a voice path that keeps working even if the AI does not.
Capture the signal
Approved PTT audio, incident ID and unit status, each with a timestamp.
Demo: synthetic voice clips played back with pre-written transcripts; no recording, no live capture.Understand speech
Speech-to-text, speaker attribution and event extraction.
Demo: transcripts are pre-written, not machine-transcribed; no speech-to-text, no speaker attribution.Check the incident
The Signal agent extracts explicit claims; Safety checks each claim against its numbered source. Production could add approved incident context and policy.
Demo: three sequential model calls in connected mode; no CAD, retrieval or agency data connected.Brief the responder
The Briefing agent drafts a short advisory update from facts that passed Safety, with links to original messages.
Demo: a scripted draft or optional model draft; a responder must verify it against the source.Extending the same pattern across the Unified Communications portfolioPTT MCVIDEO MCDATA LOCATION Scope note: this demo implements the PTT path only. The other modalities are design intent, not working code.
The three-agent contract — extract with a source ID, review each claim against its origin, draft only from approved claims — is modality-independent. The intake changes; the governance does not.
Voice transmissions
Speech-to-text per transmission; the source ID is the transmission itself. Demonstrated here.
Video events
Frame and clip events become claims with a timestamp and camera ID as the source. A responder verifies against the original clip, exactly as with audio.
Messages and forms
Text messages, status updates and structured field data are already machine-readable, so Signal extraction is simpler and Safety review is stricter about turning a request into a completed action.
Unit position and status
Position and availability are corroborating context rather than claims: they can support or contradict a stated fact, and should never originate one.
Safety contractAdvisory output only. No automatic talkgroup changes or dispatch. An operator checks the original transmission and corrects the text before a new message becomes trusted context.
GuardrailsUnclear text excluded; each candidate fact reviewed; all advisory drafts require human verification.
ObservabilityProposed: track per-agent latency, unsupported or missed facts, speech errors and operator corrections.
FallbackOriginal communications remain available when AI fails.
Customer production design.
Candidate architecture for a supervised pilot. Interface access, hosting and targets require discovery with the agency and product teams.
HLD / Major components
Independent from primary voice1 · Authorized sourcesPTT media copy, incident IDs, approved CAD and unit state
2 · Async intakeAdapter, timestamp, identity check and event queue
3 · AI processingSpeech service → Signal fact extraction → Safety fact-level gate → Briefing draft; retrieval only if approved later
4 · Briefing servicePolicy and authorization gate, claim provenance, human review, audit and versioned output
5 · Responder clientRole-scoped view, original source, correction and fallback
Media is copied for analysis. The AI path never sits in the live voice path.
LLD / Interfaces and behavior
Example contractsIngest eventagency_id, incident_id, source_id, timestamp, media_ref; deduplicate by event ID.
Speech resultTime-coded text, speaker label, confidence and model version; flag uncertain spans.
Fact recordfact_id, claim, source_id and span, confidence, approval state and expiry; reject unreadable audio.
Briefing APIAuthorize user and incident; return reviewed draft, approved fact IDs, source links, age and uncertainty.
Failure ruleTimeout or missing authorization returns no new briefing; preserve original source.
Deployment & dataChoose agency site, private cloud or approved public cloud. Define residency, retention and encryption before integration.
Quality & scaleMeasure p95 latency, word errors, harmful omissions, throughput and noisy-channel performance against agreed targets.
Release & operationsReplay test incidents, canary by agency, version models, alert on failures and rehearse rollback with an owner.
Implemented here: scripted incoming text, three agent stages (scripted or optional browser LLM calls), human review guidance, source links and visible fallback. Production proposal: authorized media adapters, event queues, speech service, context retrieval, fact-level policy gate, monitoring and agency integration.
What happens when AI fails?
The primary voice path stays independent from inference. The assistant withholds untrusted advice and retains the source.
Normal path
Communications continue through the existing service. Signal, Safety and Briefing produce an advisory draft with links to the original source messages.
Degraded path
Timeout, no approved facts, unclear speech or network loss withholds a new draft. Keep original communications and show the reason.
Security boundary
Access by role and incident. Retain source provenance and audit; apply agency deployment and retention rules.
Test the decision
Simulated controlsProposed: agency-managed inference and records. This browser demo sends synthetic text directly to Claude or OpenAI only in connected mode; there is no browser speech recognition.
What to measure
No production telemetryVoice availabilityInference latencyTranscription qualityOperator trustHarmful omission rate
No simulation active. Voice path remains independent of AI.
Operate the speech and three-agent service.
Production needs approved speech data, versioned prompts and models, fact-level evaluations, monitoring and rollback owners.
Representative data
Languages, accents, radio codecs and noisy incident conditions.
Agree on consent, retention and a human-reviewed test set.Measure failure
Speech error, missed facts, unsupported facts, Safety false approvals, source accuracy and per-agent latency.
Set release thresholds with responders and safety owners.Control change
Version speech models, agent prompts and output schemas; gate each change with replay tests and a small canary.
Compare against the previous approved version.Watch and recover
Monitor per-agent errors, quality drift, operator corrections, cost and timeouts; roll back a prompt or model if needed.
Keep an owner, on-call procedure and audit record.LEADERSHIP / DELIVERY PLAN
Small scope first. Evidence before rollout.
The architect aligns operations, product, speech/ML engineering, security and the customer around one three-agent use case: an advisory catch-up for a late responder with source verification and fallback.
Agree on the problem
Interview dispatch and field teams. Identify authorized PTT/LMR media interfaces, recording rights, incident identifiers, data boundaries and responder acceptance criteria.
Owners: Product + operationsGate: approved use case and success measures
Test with past incidents
Build the authorized speech-to-text path and three-agent handoff. Replay approved incidents and measure transcription, fact-level Safety decisions, sources and latency.
Owners: ML + domain expertsGate: agreed quality and latency thresholds
Keep humans in control
Pilot advisory drafts with a small responder group. Monitor rejected facts, corrections, latency and fallbacks; rehearse voice-path independence.
Owners: Engineering + securityGate: safety review and support readiness
Operate and improve
Version prompts and models, observe drift, train reviewers and exercise rollback. Expand only when agreed quality and safety gates pass.
Owners: Product + SRE + customerGate: measurable benefit and owner sign-off
Principal architect decision: prioritize a three-agent late-join briefing because each extracted fact can be reviewed against the original transmission. Keep dispatch and talkgroup control outside this advisory service.
AGENT PROMPT / DECISION LOGIC
View Prompt
Inspect the instruction used for this stage and the inputs handed to it.
Agent—
Mission—
Output—
Connected mode uses these same browser call instructions. Source data is inserted only when a run starts.
Independent SolTelco concept. Synthetic incidents; no Motorola integration. Optional provider connection calls a live LLM from the browser. The demo contains scripted text only. Use fictional incident data.