From Raw Telemetry to Automated Action: Building AI Pipelines That Auto-Resolve EV Charger Faults
About this session
When you operate a large, geographically distributed fleet of EV chargers, reliability stops being a facilities problem and becomes a data engineering problem. A charger that's down means a vehicle that can't charge and a route at risk — and traditional monitoring, which turns every fault into a ticket and every ticket into a truck, breaks down at scale.
This talk walks through the architecture of an AI system that moves fleet operations from reactive break-fix to proactive, autonomous remediation. We'll trace the full path — ingestion, detection, diagnosis, and action — and dig into the decisions that made each stage work.
Key ideas we’ll cover:
• Parallel detection over one big model — why multiple specialized anomaly-detection strategies, each tuned to a distinct failure category, outperform a single general detector.
• Pre-processing before diagnosis — how structuring telemetry into reasoning-ready summaries (rather than feeding raw time-series to an LLM) improves diagnostic quality.
• An LLM-powered diagnostic layer — determining root cause, severity, and recommended action from contextualized signals.
• Knowing when to act autonomously — the confidence-and-reversibility framework for deciding when AI can resolve a fault directly and when to keep a human in the loop.
• Correlation and idempotency at scale — architectural principles that prevent duplicate dispatches and phantom alerts.
Attendees will leave with a transferable blueprint for autonomous remediation that generalizes beyond EV charging to any domain built on IoT telemetry and mixed failure modes.
Speaker
Key takeaways
- How to design multiple anomaly detection methods that work together to catch different failure modes in IoT/hardware systems
- Architecture patterns for feeding pre-processed telemetry into LLM-based diagnostic agents (Amazon Bedrock) for automated decision-making
- Lessons learned scaling from detection to auto-resolution — when to let AI act autonomously and when to keep humans in the loop