# Hausa Voice Swarm
**Production Voice Agents for West African Mobile Money**
Voice-first customer support for mobile money operators serving 10M+ monthly interactions in Hausa and English. Built for 2G/3G networks and low-literacy users: short responses, confirmation-gated financial actions, sub-800ms voice-to-voice latency.
---
## The Problem
Mobile money is the dominant financial infrastructure in West Africa, yet customer support is bottlenecked by human call centers that can't scale during peak hours. Existing voice AI solutions fail here because:
- **Hausa** is a 70M+ speaker language with almost no commercial ASR/TTS support
- **Code-switching** between Hausa and English is the norm, not the exception
- **2G/3G networks** impose hard latency and bandwidth constraints
- **Low-literacy users** need concise, confirmation-gated interactions (max 2 sentences per response)
- **Financial safety** demands explicit confirmation before any mutation (transfers, bill payments)
This system handles balance checks, money transfers, bill payments, and general support entirely by voice, routing between 4 specialized agents with cost-optimized LLM selection.
---
## Architecture
```mermaid
graph TB
subgraph "Client Layer"
PHONE["Mobile Phone SIP / WebRTC"]
WEB["Web Client WebSocket"]
end
subgraph "GKE Cluster"
subgraph "Voice Pipeline (GPU)"
VP[Pipecat Pipeline]
VAD[Silero VAD]
ASR[Whisper-Hausa ASR]
TTS[Cartesia TTS]
end
subgraph "Agent Layer (CPU)"
INTENT[Intent Classifier multilingual-MiniLM]
SUP[Supervisor Router]
BAL[Balance Agent]
TXN[Transfer Agent]
BILL[Bills Agent]
GEN[General Agent]
end
subgraph "Data Layer"
REDIS[(Redis Sessions, 30m TTL)]
CRDB[(CockroachDB REGIONAL BY ROW)]
KAFKA[[Kafka / Redpanda Event Bus]]
end
subgraph "Observability"
PROM[Prometheus]
HPA[Custom-Metric HPA active_sessions]
end
end
subgraph "LLM Routing"
FLASH[Gemini Flash 70% — simple]
GPT4O[GPT-4o 30% — complex]
end
subgraph "Downstream"
ANALYTICS[Project C Analytics Pipeline]
en …