CLOUD-EDGE-IOT VOICE INTELLIGENCE
How NovoVox Works
NovoVox connects smart-meter data, structured energy intelligence and a real-time conversational AI pipeline through a secure, modular architecture.
The smartphone acts as the primary user-side edge device. Swiss-hosted services manage sessions, data, AI orchestration and evidence, while GPU workers execute the voice-processing tasks required for responsive conversations.
Energy Data
- CLEMAP API
- 10-second readings
- 1-minute readings
Data Intelligence
- PostgreSQL
- Profiles and patterns
- Tariffs and anomalies
Call Control
- Secure WebSocket
- Session orchestration
- Conversation memory
AI Pipeline
- VAD and ASR
- LLM reasoning
- Streaming TTS
User Experience
- Android voice client
- Incoming and outgoing
- Spoken guidance
Energy-data integration and intelligence
Energy Data Becomes Structured Intelligence
NovoVox ingests CLEMAP smart-meter data at 10-second and 1-minute granularities. The operational source of truth is a Django/PostgreSQL backend, not a collection of CSV files.
A scheduled processing layer aggregates phases, calculates energy and cost indicators, identifies peaks and anomalies, and builds daily, weekly and hourly profiles. Conversational agents use this structured knowledge rather than receiving raw meter records.
Raw Data
10-second data
High-frequency active and reactive power tracking.
1-minute data
Phase-level voltage, current, power and energy measurements.
Processed Profiles
01
Daily
02
Weekly
03
Hourly
04
Tariff
05
Anomaly and pattern outputs
Agent - ready
Daily profile
Weekly
Hourly
Tariff
Anomaly and pattern outputs
Energy-data integration and intelligence
A Streaming Voice Pipeline
During an interaction, the Android client streams microphone audio through a secure WebSocket connection. The Call Control Server manages the session and routes work to available GPU workers. Voice Activity Detection identifies speech boundaries, ASR creates the transcript, the LLM generates a contextual response, and TTS begins returning spoken audio in chunks.
Mobile and multilingual interaction
A Voice-First Android Experience
The validated Android application supports secure login, profile and preference loading, outgoing and backend-initiated incoming interactions, microphone capture, real-time audio streaming, spoken-response playback, call history and session management.
- English - Functionally validated.
- French - Functionally validated.
- German- Functionally validated.
- Italian - Functionally validated.
Programme and ecosystem
Secure, Modular and Scalable by Design
NovoVox separates the user interface, orchestration services, energy-data services and AI workers. Components can scale independently, supporting controlled Swiss-hosted deployment and a clear path from internal testing to multi-user pilot operation.
01
User edge
Android audio capture, playback and call interaction.
02
Swiss infrastructure
Call control, authentication, orchestration, database and monitoring.
03
AI workers
Modular GPU workers for VAD, ASR, TTS and offline post-call analysis.
04
Energy IoT
CLEMAP and Pilot 8 data interfaces provide the metering context.
Responsible AI, privacy and consent
Responsible AI and Privacy by Design
NovoVox is designed to obtain informed consent before recording and processing voice interactions. Users remain free to decline the service, and recommendations are presented as optional guidance rather than instructions.
Emotion analysis is performed offline after the interaction for quality and satisfaction assessment. It is not used to manipulate the live conversation and is not a clinical or psychological diagnosis.
01
Consent first
Recording and processing are explained before the interaction proceeds.
02
Swiss-hosted core
Core conversational and data services are hosted in Switzerland.
03
Minimal push data
Firebase Cloud Messaging carries push tokens and technical notification metadata, not sensitive conversation content.
04
Explainability
Recommendations should identify the data pattern or reason behind the suggestion.
05
No live emotional manipulation
Emotion metadata remains outside the real-time response path.