Modern vehicles generate vast telematics data streams, yet unexpected breakdowns persist due to limitations in real-time insight and centralized analytics. OEMs also struggle with inconsistent maintenance outcomes and privacy hurdles.
This platform introduces a breakthrough: A privacy-preserving, agentic AI ecosystem for predictive maintenance, powered by Federated Learning and multi-agent orchestration.
- Federated Learning: Collaborative model training at the edge—no raw data leaves the vehicle!
- Agentic AI System: Autonomous agents diagnose, forecast, communicate, schedule, and secure.
- Gemini LLM Integration: Human-readable explanations tailored for vehicle owners.
- Interactive Streamlit Dashboard: Real-time simulation, insights, and recommendations.
This repository showcases the working prototype submitted for Zone-2, EY Techathon 6.0.
- Local nodes simulate decentralized vehicle data and train ML models.
- Global model aggregated via FedAvg; performance tracked across rounds.
- AUC chart visualizes learning progress.
Orchestrated by a Master Agent, with dedicated Worker Agents for:
- Telemetry Analysis
- Failure Prediction
- Owner Communication (Gemini-LangChain)
- Demand Forecasting
- Service Scheduling
- Lifecycle Management
- RCA/CAPA Insights
- UEBA Security Surveillance
- Sliders for telematics input simulation
- Instant failure probability & classification
- Diagnostics: Engine, Cooling, Suspension, Usage Fatigue
- Trends: Last 10 simulated trips
- Federated Learning progress chart
- Gemini-powered recommendations
- Raw data remains local—privacy by design
- Continuous behavioral monitoring via UEBA agent
- Modular, scalable architecture ready for expansion
- Vehicle Nodes (FL Clients): Local training on synthetic telemetry
- FL Server: Aggregates weights → global model
- Master Orchestrator: Central controller of agents
- Worker Agents: Specialized tasks per system state
- Owner Interface: Concept for notifications (voice/mobile)
- Dashboards: Maintenance, forecasting, RCA
- (Architecture diagram available in PPT submission)
| Layer | Tools / Technologies |
|---|---|
| Programming | Python 3.10 |
| ML Frameworks | Scikit-learn, Flower (FL) |
| Dashboard | Streamlit |
| LLM | Google Gemini Pro |
| Data | Synthetic telemetry simulator |
| Visualization | Matplotlib, Streamlit charts |
| Security | UEBA conceptual framework |
| Deployment | Local / microservice-ready |
├── data/
│ └── simulated_vehicle_data.csv
├── models/
│ ├── global_model.pkl
│ └── fl_metrics.pkl
├── generate_data.py
├── federated_training.py
├── streamlit_app.py
├── README.md
├── requirements.txtFor any queries or collaborations, reach out to Akash Karri.
- Prototype submitted for EY Techathon 6.0, Zone-2.
- Expand agentic architecture for EVs and commercial fleets.
- Integrate advanced anomaly detection and reinforcement learning.
- Real-world pilot with OEM partners.