Back to Projects
AIIndustrialEU R&DDockerIoT

AVAROS — Voice-First AI for Sustainable Manufacturing

Leading the architecture and development of AVAROS — a voice-first AI assistant for sustainable manufacturing. Features OVOS voice AI, DocuBoT RAG knowledge assistant, PREVENTION predictive maintenance, platform-agnostic adapter pattern, and Clean Architecture with 7 documented design decisions. Targets 8% electricity reduction, 5% materials efficiency, and 10% CO₂ reduction under the EU WASABI consortium.

Role: Lead Developer & System ArchitectPeriod: 2025–2026Funding: EU — WASABI Consortium
View Code

8%

electricity reduction target

5%

materials efficiency target

10%

co2 reduction target

Auto1 / 4
AVAROS — Voice-First AI for Sustainable ManufacturingAVAROS — Voice-First AI for Sustainable Manufacturing
AVAROS — Voice-First AI for Sustainable ManufacturingAVAROS — Voice-First AI for Sustainable Manufacturing
AVAROS — Voice-First AI for Sustainable ManufacturingAVAROS — Voice-First AI for Sustainable Manufacturing
AVAROS — Voice-First AI for Sustainable ManufacturingAVAROS — Voice-First AI for Sustainable Manufacturing

Problem

Manufacturing operators need real-time access to energy KPIs, anomaly alerts, and sustainability metrics — but they can't stop working to navigate dashboards or read reports. Existing EnMS platforms are locked to specific vendors, require technical expertise, and offer zero conversational intelligence. Factory floors need a voice-first AI that operators can query hands-free while working, backed by predictive maintenance and RAG-powered document retrieval.

My Role & Responsibilities

  • Lead Developer and System Architect — primary technical leadership across the full AVAROS system
  • Designed the Clean Architecture with 7 documented design decisions (DEC-001 to DEC-007) and strict SOLID compliance
  • Built the platform-agnostic adapter pattern (DEC-001) — AVAROS works with RENERYO, EnergySuite, GreenMetrics, or ANY manufacturing platform without code changes
  • Implemented the Universal Metric Framework (DEC-002) — canonical manufacturing concepts mapped across platforms
  • Developed the OVOS voice skill with 5 query types: KPI retrieval, comparison, trend analysis, anomaly detection, and what-if simulation
  • Integrated DocuBoT — RAG-based knowledge assistant for procedures, specs, and compliance documents
  • Integrated PREVENTION — anomaly detection and predictive maintenance module
  • Built the FastAPI + React web dashboard with settings wizard, metric mapping, and API key auth
  • Managed Docker-Compose orchestration for the full stack with Nginx TLS reverse proxy
  • Supervised a junior developer — code reviews, task delegation, architectural mentoring
  • Implemented GDPR-compliant audit trails, RBAC, data minimization, and security hardening

Platform Screenshots

Architecture

Built on Clean Architecture with strict layer separation — domain never imports infrastructure:

  • Presentation Layer: OVOS voice interface + React/TypeScript dashboard
  • Application Layer: FastAPI orchestration and query routing (KPI retrieval, comparison, trend, anomaly, what-if)
  • Domain Layer: canonical metric model + immutable business entities
  • Integration Layer: platform adapters (RENERYO, EnergySuite, GreenMetrics, Mock)
  • Data/Infra Layer: TimescaleDB, MQTT/Modbus, Docker-Compose, Nginx TLS, RBAC/audit trail
OVOS / Web UI
      ↓
FastAPI Orchestrator
      ↓
Domain Use-Cases + Metric Model
      ↓
Platform Adapters
      ↓
TimescaleDB + MQTT/Modbus + Secure Runtime

7 Design Decisions

  • DEC-001 — Platform-agnostic: works with any manufacturing platform via adapter pattern
  • DEC-002 — Universal Metric Framework: canonical metric names mapped per platform
  • DEC-003 — Clean Architecture: strict layer separation; domain never imports infrastructure
  • DEC-004 — Immutable Domain Models: frozen dataclasses for thread safety and predictable behavior
  • DEC-005 — Zero-Config First Run: MockAdapter fallback; docker compose up gives a working system
  • DEC-006 — Settings Service Pattern: no hardcoded credentials in application code
  • DEC-007 — Intelligence in Orchestration: adapters fetch data; intelligence lives in QueryDispatcher, DocuBoT, and PREVENTION

Tech Stack

  • Voice AI: OVOS (Open Voice OS) — custom skill with domain-specific intent handling and 5 query types
  • Knowledge AI: DocuBoT — RAG-based document retrieval for procedures and compliance specs
  • Predictive Maintenance: PREVENTION module — anomaly detection, predictive alerts, and equipment health scoring
  • Backend: FastAPI (Python) — Clean Architecture with domain/use case/adapter/presentation layers
  • Frontend: React + TypeScript + Vite — settings wizard, metric mapping UI, dashboard
  • Database: TimescaleDB (PostgreSQL with time-series extensions)
  • Infrastructure: Docker-Compose, Nginx (TLS 1.2+, security headers, Let's Encrypt support)
  • IoT: MQTT, Modbus TCP sensor integration
  • Security: RBAC, GDPR audit trails, data minimization, API key auth
  • Code Quality: SOLID compliance with documented examples; test pyramid (60% unit, 30% integration, 10% E2E); domain 100% coverage target; functions max 20 lines, files max 300 lines
  • Language Mix: Python 65.9%, TypeScript 30.0%

Results & Impact

The AVAROS system targets measurable sustainability improvements validated through pilot deployments:

  • ~8% electricity reduction through intelligent monitoring and voice-accessible optimization recommendations
  • ~5% materials efficiency improvement via predictive analytics and waste reduction alerts
  • ~10% CO₂ reduction aligned with EU sustainability goals
  • Platform-agnostic by design — the adapter pattern means AVAROS can plug into any manufacturing platform (RENERYO, EnergySuite, GreenMetrics, or custom) without changing a single line of business logic
  • Voice-first factory access — operators query KPIs, comparisons, trends, anomalies, and run what-if simulations hands-free while on the production floor
  • Zero-config deployment — docker compose up → fully operational system with mock data for immediate demonstration

Challenges & Lessons Learned

  • Integrating four independent subsystems (OVOS, DocuBoT, PREVENTION, RENERYO) into a cohesive product required rigorous API contract design, version negotiation, and the DEC-007 principle that intelligence lives in orchestration, not adapters
  • Clean Architecture enforcement — maintaining strict layer boundaries across a growing team required documented design decisions, automated linting rules, and regular architecture reviews
  • Leading a junior developer under EU project delivery pressure taught me the value of clear task decomposition, pair programming on critical paths, and regular code review cadence
  • Docker-Compose orchestration for 8+ services with TLS, health checks, volume management, and graceful degradation across dev/staging/production environments

How AI/Agents Were Used

AVAROS is AI-native in three dimensions: (1) the product IS an AI voice assistant with RAG retrieval and predictive intelligence; (2) I orchestrated agentic development workflows using GitHub Copilot Pro+ with custom VS Code agents — they generated architecture proposals, scaffold code, and test suites that I validated and hardened for EU project delivery standards; (3) the Clean Architecture itself was co-designed with AI agents, then pressure-tested against SOLID principles and real manufacturing use cases.