FixSA Voice

The problem

Residents describe reality. Systems demand fields.

Phone queues, fragmented forms, missing landmarks, repeated reports, and opaque follow-up all waste time. Voice can close that gap—if the product preserves consent, structure, safety, and operator control.

Product principles

Ask only for information needed to act.
Explain duplicates instead of silently merging.
Never create before explicit confirmation.
Stop ordinary automation when danger is immediate.
Keep private data out of public views.
Say clearly what is demo, synthetic, or disconnected.

Architecture

One contract from microphone to work order.

1. Voice edge

AssemblyAI Voice Agent streams English transcripts, spoken replies, and tool calls using short-lived server-issued tokens.

2. Civic orchestrator

Zod schemas validate classification, confirmation, duplicate, evidence, status, priority, and SLA operations.

3. Durable work orders

Laravel and MariaDB provide transactional writes, public/private projections, audit history, idempotency, and indexed queues.

4. Reliable judge mode

The same experience remains demonstrable through synthetic browser-local data when no API is configured.

Privacy by design

Data minimisation is visible

Consent appears before recording. Audio is ephemeral in the demo. Public tracking excludes contact, identity, precise-home, transcript, internal-note, and assignment detail.

Limitations

This is not municipal dispatch

Locations and work orders are synthetic, routing teams are fictional, metrics are simulated, and no record is dispatched to a municipality. Emergency help remains the responsibility of official services.

Roadmap

From demo to authorised service

Validate locally

Accent testing, category vocabulary, field policy, SLA rules, and accessibility research with consenting participants.

Integrate safely

Authorised work-order API, identity and abuse controls, geocoding, retention policy, data processing terms, and operator training.

Expand responsibly

Only expose languages actually supported and tested for the selected model; measure equitable completion and transcription quality.

Civic intake that starts with a human voice.