Chat With Your Data — Kubernetes Conversion

Project Overview

A working AI chatbot on infrastructure nobody else could safely operate is a liability wearing a demo. This project converted a production "chat with your data" chatbot from bespoke, hand-managed infrastructure to Kubernetes-native infrastructure as code.

The Challenge

  • Tribal Knowledge Infrastructure: Deployment was a checklist one person knew, not a repeatable process
  • Scattered Secrets: Configuration and credentials lived in more than one place, with no single source of truth
  • No Real Observability: There was no reliable way to answer "is this degraded right now" beyond someone noticing
  • Bus-Factor Risk: The system's operational knowledge lived in one person's head

Technical Solution

Helm-Packaged Deployment

The chatbot's deployment definition moved into a Helm chart — a single, versioned, source-controlled description of what "correctly deployed" means, instead of a sequence of manual steps.

GitOps Deployment Pipeline

Deployments now happen through a GitOps controller reconciling the cluster against the Helm chart in source control. A deployment is a merged pull request with a visible diff, not a person running kubectl apply from memory.

Centralized Secrets and Real Observability

Secrets moved into a centralized, managed store instead of being scattered across configuration files and environment variables. An observability stack was added so the team has metrics, logs, and alerting instead of finding out about degradation from a user complaint.

Results and Impact

  • Deployment is now a repeatable, auditable process anyone on the team can follow
  • Secrets live in one managed place instead of several
  • The team has real visibility into the chatbot's health instead of relying on someone noticing something's wrong

Key Learnings

Migrating working software off bespoke infrastructure is unglamorous work that rarely shows up as a headline feature, but it's the difference between a system the team can operate and a system that depends on one person staying available. GitOps and Helm didn't change what the chatbot does — they changed who can safely operate it.