SignalForge
Complete technical architecture · Voice AI Incident Commander

From live voice to verified operational outcomes.

Agora supplies the conversational media plane. An event-sourced control plane builds authoritative incident state. A bounded specialist-agent council proposes updates. Deterministic policies protect critical actions, while outcome monitors prove whether mitigation actually worked.

Sub-second participationEvidence-linked beliefsHuman approvalExecution ≠ recoveryAuditable replay
01 · End-to-end system

The complete architecture

Select a layer to inspect how voice and telemetry become a validated incident-state mutation rather than an unchecked model response.

Participants

Incident CommanderAuthority and priorities
EngineeringEvidence and investigation
Support & BusinessImpact and communication
SignalForge voiceClarifies and summarizes
1

Agora real-time voice plane

Interruptible room participation, speaker identity, VAD and committed transcripts.

Agora RTCConversational AISpeaker mappingBarge-in
ordered incident event stream · Redis Streams
2

Ingestion and normalization

FastAPI/WebSocket commits utterances, monitoring events and provenance.

Event schemaDeduplicationRole resolverIdempotency
3

Deterministic incident-state engine

Versioned facts, hypotheses, evidence, decisions, missions, actions and risks.

State reducerEvidence graphConflict engineTimeline
4

Bounded specialist-agent council

Structured proposals for extraction, skepticism, risk and next investigation.

ExtractorEvidence criticRisk guardianPlannerComms
5

Policy, approval and action gateway

Risk-to-authority rules, explicit confirmation and decision contracts.

RBAC/ABACApprovalAuditTool sandbox
6

Outcome verification and reflection

Metric windows prove verified, failed or inconclusive; failures trigger replanning.

Metric watchersSuccess criteriaDynamic replanSummary policy

Integrations

MonitoringDatadog, Grafana, Prometheus
CoordinationSlack, Jira, Statuspage
On-callPagerDuty schedules
OperationsRunbooks, deploys, rollback
KnowledgePostmortems and ownership
LAYER 1

Agora real-time voice plane

Maintains a low-latency, interruptible room experience. Only committed utterances enter the event stream, keeping partial ASR revisions out of authoritative state.

02 · Runtime behavior

One utterance, safely processed

Step through the actual message path from speech to validated state, tool query, spoken response and verification.

Ready: APAC checkout failures are 38% enters the live room.
03 · Agentic intelligence + safety

Agents propose. The system decides.

Models handle ambiguity; deterministic code owns truth, permissions, state transitions and auditability.

Orchestrator
shared snapshot
typed proposals only
Extraction agentClaims, roles, decisions, actions.
Evidence criticContradictions and missing scope.
Risk guardianCustomer and operational risk.
Investigation plannerHighest-value uncertainty check.
Communication agentConcise role-aware summaries.
Decision contract

Critical action boundary

Rollback payment-service 4.17HIGH RISK
Evidence
Deployment correlation; alternative cause unresolved
Authority
Incident Commander or Operations Lead
Success
5xx < 2%; checkout > 97% for 5 minutes
Reversal
Abort if reconciliation lag exceeds 90 seconds
Agent proposal
→
Policy + human confirmation
→
Idempotent executor
Blocked by default: root-cause declarations, external status updates, pages and production mutations without required authority.
04 · Truth and memory

The incident is a typed state machine

PostgreSQL is authoritative. Agents receive a versioned snapshot and return a validated proposal; they never write truth directly.

Incidentseverity
phase
commander
state_version
Claimtype
status
confidence
speaker_id
Evidencesource
timestamp
scope
reliability
Edgesupports
contradicts
supersedes
Decisionrationale
approver
success
reversal
Missionowner
evidence_needed
deadline
Actioncapability
risk
approval
result
Riskimpact
likelihood
mitigation
State invariants
✓A hypothesis needs linked evidence and authorized confirmation to become fact.
✓Every action has one accountable owner, deadline and terminal status.
✓Executed and verified effective are separate states.
✓Conflicting claims coexist until evidence resolves scope.
✓Every mutation records event, proposal, validator and state version.
✓Immutable event replay rebuilds the same state.
05 · Build and deployment

Hackathon-feasible, production-shaped

One payment scenario, one monitoring adapter and one workflow integration prove the complete loop.

Browser / incident room

React + TypeScriptEvidence, timeline, approvals
Agora Web SDKAudio and participant events
WebSocketAuthoritative state diffs

Application + intelligence

FastAPI gatewayAuth, incidents, events
Redis Streams workersOrdering, retries, replay
Agent orchestratorStructured output and budgets
Policy/action serviceApprovals and idempotency
Verification schedulerOutcome windows and replanning

Data + external systems

PostgreSQLTruth and audit
pgvectorRunbooks and similar incidents
Object storageRetention-bound artifacts
AdaptersMonitoring, Slack/Jira, PagerDuty
OIDC + tenant isolationSecrets vaultEncryptionPII redactionLeast privilegeOpenTelemetryModel versioningDead-letter replay
06 · Guided proof

Run the payment-outage simulation

Watch SignalForge weaken a premature conclusion, protect a critical action and verify the outcome.

Live evidence graph

Release 4.17 caused the outage

72% hypothesis
Current belief: plausible
SIGNALFORGE SPOKEN INTERVENTIONI have one regional fact and an unverified deployment hypothesis. I will not declare root cause. Next: compare gateway throttling and checkout success by region.