An AI assistant that answers IT troubleshooting questions from a team's own runbooks, safely, with sources, on Microsoft Azure.
Rapid-fail protection is likely stopping the pool after repeated crashes [1]
5002 / 5011w3wp.exe [2]iisreset restarts every site on the server.
Operations teams keep their hard-won knowledge in runbooks: how to fix a crashing IIS application pool, an unreachable EC2 instance or a failing deployment pipeline. During an incident, finding the right page and the right step takes time, and generic AI chatbots don't know your environment.
Worse, they can invent commands that sound plausible but are wrong.
It finds the most relevant sections, checks that they really answer the question, and replies with step-by-step guidance and the sources it used. If the runbooks don't cover a question, it says so instead of guessing.
Around the AI sits everything a business needs to trust it in production: guardrails, automated quality testing, monitoring and a secure, fully scripted cloud deployment.
Every question goes through the same four steps, and each step is traced so the team can see exactly why an answer was given.
A data pipeline validates the runbooks, splits them by section and indexes them for both keyword and meaning-based search. Only changed files are re-processed.
Guardrails check the question first. Then the assistant works out the topic (networking, CI/CD, security and more) to narrow the search.
Hybrid search finds the best-matching sections. A grader checks they're relevant; if not, the question is rephrased and searched again.
The reply gives clear steps, cites the runbook sections it used and warns about risky commands before anyone runs them.
Answers come only from your documentation, with the sources shown, so engineers can check every step.
Secrets such as passwords and keys are masked, prompt-injection attempts are blocked, harmful content is filtered and usage is rate-limited.
Commands that can cause outages, such as server restarts or deletions, are highlighted with a clear warning.
If the first search misses, the assistant rephrases the question and tries again before answering.
An evaluation suite of real and adversarial questions scores every version, so improvements are proven, not assumed.
Runbooks are validated before indexing, and updates are incremental, idempotent and safe to re-run.
The whole environment is defined in Terraform and deployed with scripted, reviewable steps. No passwords or API keys live in the code or the container image.
All of it provisioned with Terraform, with remote state protected by Entra ID.
Runbooks, policies, product manuals or support articles: we can build a secure assistant on your own documentation and your cloud.