Nexus Apollo

Production edge AI: Raspberry Pi 5 + Hailo-8 NPU + 1TB SSD monitoring live 460V commercial HVAC. 21+ sensors, 8-model ensemble, TiDB vector search, PostgreSQL & SQLite for fault detection.

Comment

Inspiration

HVAC systems generate massive amounts of sensor data that contains patterns indicating potential failures. Traditional threshold-based monitoring misses complex multi-sensor correlations. TiDB's vector search capability offered a way to find similar fault patterns in historical data, enabling predictive maintenance.

What it does

Nexus Apollo is an HVAC monitoring and diagnostic system that:

  • Collects real-time sensor data (temperature, pressure, humidity, electrical) from HVAC equipment
  • Converts sensor readings into vector embeddings stored in TiDB Cloud
  • Uses vector similarity search to match current conditions against known fault patterns
  • Runs 8 specialized AI models for different HVAC subsystems (electrical, refrigeration, airflow, etc.)
  • Provides a web dashboard for monitoring and running diagnostic workflows

How we built it

  • Backend: Node.js/Express API with PostgreSQL for customer/equipment data, SQLite for sensor readings, and TiDB Cloud for vector storage
  • Frontend: Next.js 14 with TypeScript and shadcn/ui components
  • Vector Search: TiDB Cloud storing 1536-dimensional embeddings with HNSW indexing
  • AI Models: 8 ONNX models running inference for specialized fault detection
  • Multi-Step Workflow:
    1. Collect sensor data
    2. Generate embeddings
    3. Search TiDB for similar patterns
    4. Run specialized AI models
    5. Aggregate results
    6. Generate recommendations

Challenges we ran into

  • Connecting to TiDB Cloud from Node.js required proper SSL certificate configuration
  • Creating proper vector embeddings from time-series sensor data
  • Coordinating multiple AI models in a sequential workflow
  • GitHub's 100MB file size limit when pushing the repository

Accomplishments that we're proud of

  • Successfully integrated TiDB vector search for pattern matching
  • Built a complete 6-step agentic workflow exceeding the 3-step requirement
  • Created a working dashboard with real-time data visualization
  • Implemented equipment-specific diagnostic workflows

What we learned

  • TiDB's vector search is powerful for finding similar patterns in sensor data
  • Multi-model approaches provide better fault detection than single models
  • Proper indexing is critical for vector search performance

What's next for Nexus Apollo

  • Expand to more equipment types
  • Improve vector embedding generation
  • Add more sophisticated fault prediction algorithms
  • Deploy to production facilities

Built With

Share this project:

Updates