Skip to main content
Shreemanto.dev

Command Palette

Search for a command to run...

AI Engineering

At Okta I build production AI systems for infrastructure and ops. The work below is grounded in what I actually shipped: Infra AI, The Guardian, and Terraform plan intelligence on Amazon AgentCore, MCP, and AWS Bedrock.

Some implementation details are left out on purpose. The systems, integrations, and outcomes match the work on my resume.

AI Stack

AI Platform

Amazon AgentCoreAWS BedrockMCPAI AgentsLLM Integration

Integrations

SpinnakerKubernetesTerraformPagerDutyAWS

Ops & Knowledge

Amazon KendraConfluenceSlackGrafanaDatadog

Runtime

AgentCore RuntimeAgent GatewayLambda → AgentCore migration

Techniques

Semantic chunkingHierarchical summarizationConversational memorySession management

Outcomes

22+ AWS environments1,000+ users1,000+ release trains/day

Infra AI

Production

Internal AI platform that integrates infrastructure systems, operational knowledge, and production tooling so engineers can investigate production issues through contextual infrastructure search across 22+ AWS production environments. Built with Amazon AgentCore and MCP, then scaled from 50 to 1,000+ users.

Architecture

  • Production AI agents on Amazon AgentCore for contextual infrastructure search and analysis
  • MCP servers for Spinnaker, AWS, Kubernetes, Terraform, Grafana, Datadog, and PagerDuty
  • Conversational memory, session management, and UX improvements for multi-turn investigations
  • Migrated agent infrastructure from AWS Lambda + ALB to Amazon AgentCore Runtime and Agent Gateway
  • Supports platform and SRE workflows across 18+ production environments

Lessons Learned

  • Tool design matters as much as model choice. Vague tools get you hallucinated actions.
  • Infrastructure agents need guardrails. Prefer reads before writes.
  • Memory and session UX are what turn a demo into something people actually use.
Amazon AgentCoreMCPAWS BedrockSpinnakerKubernetesTerraformGrafanaDatadogPagerDutyAWS

The Guardian

Production

Internal agentic AIOps assistant for on-call engineers. Integrates Slack, Confluence, Amazon Kendra, and platform services to provide contextual incident assistance without leaving the incident workflow.

Architecture

  • Agentic AIOps assistant wired into Slack for on-call workflows
  • Operational knowledge retrieval through Confluence and Amazon Kendra
  • Platform service integrations for live infrastructure context
  • Designed to give contextual incident assistance at the moment an engineer needs it

Lessons Learned

  • For ops tools, retrieval quality beats raw model quality.
  • People trust the system more when answers show up where the work already happens (Slack).
  • Contextual assistance only sticks if the first answer is grounded and useful.
Amazon AgentCoreAWS BedrockAmazon KendraSlackConfluenceMCPPython

Terraform Plan Intelligence

Production

AI-driven Terraform plan intelligence on AWS Bedrock. Uses semantic chunking and hierarchical summarization to generate deployment summaries and risk assessments, scaling to 1,000+ release trains/day with plans that can contain tens of thousands of resource changes.

Architecture

  • AWS Bedrock for deployment summaries and risk assessments
  • Semantic chunking for large Terraform plans
  • Hierarchical summarization to keep signal as plan size grows
  • Built for release-train volume, not one-off PR demos

Lessons Learned

  • Huge Terraform plans need structure before an LLM can be useful.
  • Risk summaries matter more than perfect prose for release reviewers.
  • Fit the output into existing release workflows or nobody adopts it.
AWS BedrockTerraformPythonSpinnakerAWS