Incidents that explain themselves

AI-Ops case study, 2025 to 2026. Role: Built and led.

Heartbeat AI-OPS on an in-house LLM, then guardrailed AI agents that let any engineer run a root-cause analysis.

  • $44K saved per month
  • 38 guardrailed MCP tools
  • 3 agent plugins
  • #1 Hackweek 2025

The problem

Signals were scattered: alerts in Zenduty, logs and metrics in different tools, deployments and DB changes elsewhere. On-call engineers spent their time stitching context together instead of fixing things, and nobody could see idle spend or deployment drift at a glance.

How it works

One engine for scattered signals

Heartbeat connects Zenduty, application logging, GCP, app deployments, DB changes and GKE into a single cohesive engine.

An LLM that stays in-house

An in-house Mistral model analyses the stream, producing DORA metrics, alert-noise insights and deployment deviations, all from one place.

Pay only for what runs

Removing idle resources and consolidating tools saved $44K a month. Heartbeat won Blibli Hackweek 2025 and a Spot Award.

RCA for every engineer

As lead, Kirat shipped agent plugins (DevOps RCA, Datadog, SigNoz) to the team marketplace: 38 guardrailed MCP tools that turn an incident into a Confluence-ready RCA, with Jenkins automation.

Outcome

  • $44K a month saved, recognised with the 2025 hackathon win and a Spot Award.
  • Any engineer can run a guided root-cause analysis without being an observability expert.
  • RCA write-ups are generated consistently, ready for Confluence.

This initiative, built entirely in-house from the ground up, establishes a new standard for infrastructure monitoring and optimization within our organization.

Satish Srinivas, Senior DevOps Manager ยท Spot Award, Oct 2025

Stack: Mistral (in-house), MCP, Datadog, SigNoz, Zenduty, Jenkins, Confluence.

Next case study: GCP to Tencent Cloud, secrets first