Barun Acharya

AIOps powered by Large Language Models is everywhere—and so is the bill.

As organizations rush to plug LLMs into incident response workflows, they're discovering an uncomfortable reality: operational data is massive, noisy, and expensive to reason about. More context often means more tokens, more latency, and more cost.

Meanwhile, many teams are still debating whether OpenTelemetry is worth the effort.

This talk explores an unexpected connection between the two. Through a series of experiments on real Kubernetes incidents, we compare traditional log-analysis techniques, modern LLM-based approaches, and emerging agentic systems. The results challenge some common assumptions about how AI should consume operational data—and reveal why a seemingly mundane observability feature may be one of the most effective ways to improve both accuracy and cost.

If you're evaluating OpenTelemetry, building AI-powered operational tooling, or simply trying to understand where AIOps is headed, this session offers a data-backed perspective on a question few teams are asking: what if the best optimization for AI isn't a better model, but better telemetry?