LLM Telemetry Copilot for Fleet Analytics

Goal

Enable non-ML stakeholders to query trip and fleet behavior without writing SQL or navigating dashboards.

Implementation

I built a retrieval-augmented agent pipeline with role-specific prompts and context-aware query planning.

• Modeled task routing with LangGraph state transitions for question classification and execution strategy.
• Added retrieval over telemetry summaries and key metrics for grounded responses.
• Deployed on on-prem GPU inference endpoints to control cost and latency.

Impact

The assistant became a practical internal tool for exploratory analysis and rapid triage.

Future Extensions

• Add feedback loops for answer-quality scoring.
• Integrate chart generation for trend-heavy questions.
• Support deeper comparative analytics between driver cohorts.