<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ginny Ibe on Medium]]></title>
        <description><![CDATA[Stories by Ginny Ibe on Medium]]></description>
        <link>https://medium.com/@ginnyibe?source=rss-c77eb2336be8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ylediQjPxy26nE1bMwJPsw.jpeg</url>
            <title>Stories by Ginny Ibe on Medium</title>
            <link>https://medium.com/@ginnyibe?source=rss-c77eb2336be8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 17:42:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ginnyibe/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When I Deliberately Created Terraform Drift to Test My Cloud Safety Net]]></title>
            <link>https://medium.com/@ginnyibe/when-i-deliberately-created-terraform-drift-to-test-my-cloud-safety-net-a4898bdf8e35?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4898bdf8e35</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[terraform]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 01:57:32 GMT</pubDate>
            <atom:updated>2026-09-17T01:57:32.870Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What a controlled Azure change taught me about Terraform, policy checks, Claude Code, and human judgment.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p6QHuRZ0l53K8ewMDQrWFg.png" /></figure><p>There is something uncomfortable about deliberately changing infrastructure outside your infrastructure-as-code workflow.</p><p>That discomfort was exactly what I wanted.</p><p>While working on my <strong>Book Review App capstone on Azure</strong>, I built a Terraform drift-and-policy review workflow. But rather than simply running it against a healthy environment, I wanted to answer a more practical question:</p><blockquote><strong>If something changes behind Terraform’s back, will my safety controls actually catch it?</strong></blockquote><p>So I created controlled drift.</p><p>And what I learned was less about detecting a changed tag and more about <strong>how automation, AI, and human judgment should work together.</strong></p><h3>☁️ Why I Wanted to Test for Drift</h3><p>Terraform gives infrastructure teams a way to define infrastructure as code and compare the desired configuration with the actual environment.</p><p>But infrastructure doesn’t always change through Terraform.</p><p>Someone might:</p><ul><li>Make a change directly in the Azure Portal</li><li>Modify a resource manually during troubleshooting</li><li>Change a configuration during an emergency</li><li>Forget to update the Terraform configuration after a manual change</li></ul><p>That creates <strong>configuration drift</strong>.</p><p>The infrastructure may still be running, but the live environment no longer completely matches what the code describes.</p><p>That’s where my experiment began.</p><h3>🧪 Creating Controlled Drift</h3><p>I went into the <strong>Azure Portal</strong> and manually added a tag to one of my Azure resources.</p><ul><li>No Terraform.</li><li>No pull request.</li><li>No code change.</li></ul><p>Just a controlled change made directly in the cloud environment.</p><p>The goal wasn’t to break the application.</p><p>The goal was to create a small, intentional difference between:</p><p><strong>Terraform configuration</strong><br> ⬇️<br> <strong>Actual Azure infrastructure</strong></p><p>Then I wanted to see whether my workflow could identify it.</p><h3>🏗️ The Three-Layer Safety Workflow</h3><p>The workflow I built has three distinct responsibilities.</p><p>I deliberately separated them because I didn’t want one component doing everything.</p><h3>1️⃣ Bash: Gather the Evidence</h3><p>The first layer is a Bash script:</p><pre>tf-drift-check.sh</pre><p>Its job is straightforward.</p><p>It runs:</p><p>terraform plan</p><p>Then exports the resulting plan into JSON and performs deterministic checks for things such as:</p><ul><li>Destructive actions</li><li>Risky ingress rules</li><li>Terraform plan status</li></ul><p>The important distinction is that the script isn’t trying to interpret the entire situation.</p><p>It gathers evidence.</p><p>📋 No opinions</p><p>No guessing</p><p>Just repeatable checks</p><p>This creates a predictable foundation for the next layer.</p><h3>2️⃣ 🤖 Claude Code: Analyze the Evidence</h3><p>The second layer is my Claude Code Skill:</p><pre>/tf-drift-review</pre><p>Instead of giving the AI unrestricted access to change infrastructure, I designed this part of the workflow around <strong>review and analysis</strong>.</p><p>Claude reads the evidence produced by the Terraform workflow and helps answer questions such as:</p><ul><li>What changed?</li><li>Which findings actually represent risk?</li><li>Which findings require additional context?</li><li>What deserves human attention?</li></ul><p>This distinction became extremely important during testing.</p><p>The automated checks produced a <strong>FAIL</strong> result.</p><p>But FAIL didn’t automatically mean:</p><blockquote><em>“Everything here is dangerous.”</em></blockquote><p>It meant:</p><blockquote><strong><em>“Stop and investigate.”</em></strong></blockquote><p>That’s a very different thing.</p><h3>🛡️ 3️⃣ PreToolUse Hook: Enforce the Boundary</h3><p>The third layer is the<em> PreToolUse</em> hook.</p><p>This is where the workflow moves from analysis to enforcement.</p><p>The rule is simple:</p><h3>🚦 FAIL → STOP</h3><p>If the latest drift report says <strong>FAIL</strong>, the hook blocks:</p><pre>terraform apply</pre><p>The hook doesn’t try to reason about whether the failure is justified.</p><p>It doesn’t debate with the AI.</p><p>It doesn’t make the final engineering decision.</p><p>It simply enforces the guardrail.</p><p>That separation matters.</p><p><strong>The AI analyzes.<br> The hook enforces.<br> The human decides.</strong></p><h3>🔎 What Happened When I Tested It?</h3><p>The Terraform plan detected my manually created drift.</p><p>The plan returned <strong>exit code 2</strong>, and the JSON output identified the resource and attribute that had changed.</p><p>So far, so good.</p><p>But then I saw something more interesting.</p><p>The overall policy result was:</p><p>🔴 <strong>FAIL</strong></p><p>My immediate reaction was:</p><blockquote>“Uh oh.”</blockquote><p>But instead of reacting to the label, I went back to the evidence.</p><p>And that’s where things became interesting.</p><h3>🧩 FAIL Doesn’t Always Mean “Everything Is Dangerous”</h3><p>The FAIL result contained findings that needed different levels of interpretation.</p><p>Several rules appeared to indicate open internet access.</p><p>At first glance, that sounds serious.</p><p>But after examining the actual rules, some were harmless <strong>outbound/deny rules</strong> rather than dangerous inbound access.</p><p>There was also <strong>one destructive change that genuinely required scrutiny.</strong></p><p>This was probably the most valuable part of the entire exercise.</p><p>A simple status such as:</p><p>🟢 PASS<br> or<br> 🔴 FAIL</p><p>doesn’t tell the whole story.</p><p>It tells you whether the workflow found something that requires attention.</p><p><strong>The evidence tells you what actually happened.</strong></p><h3>👤 Why I Still Made the Final Decision</h3><p>After reviewing the findings, I ran every terraform apply myself.</p><p>That wasn’t because AI analysis has no value.</p><p>Quite the opposite.</p><p>The AI helped make the evidence easier to understand.</p><p>But the decision to change infrastructure remained mine.</p><p>That’s an important distinction when working with agentic AI.</p><p>Giving an AI access to infrastructure doesn’t mean the AI should automatically become the decision-maker.</p><p>Instead, I wanted the workflow to create <strong>layers of control</strong>:</p><h3>🤖 AI</h3><p>Helps understand the evidence.</p><h3>⚙️ Automation</h3><p>Performs deterministic checks.</p><h3>🛡️ Guardrails</h3><p>Prevent an unsafe action from proceeding automatically.</p><h3>👤 Human</h3><p>Reviews the evidence and owns the final decision.</p><p>This is a much more deliberate approach than simply asking an AI agent to “check the Terraform and apply the changes.”</p><h3>💡 The Bigger Lesson</h3><p>Before this exercise, it would have been easy to think of terraform plan as the step you run before terraform apply.</p><p>But this experiment changed how I look at it.</p><p>The plan isn’t just a technical preview.</p><p>It is an <strong>evidence layer for an infrastructure decision</strong>.</p><p>A plan can show:</p><ul><li>What Terraform thinks will change</li><li>What resources are affected</li><li>Whether infrastructure has drifted</li><li>Whether actions may be destructive</li><li>Which findings require further investigation</li></ul><p>And when automation is involved, that evidence becomes even more important.</p><p>Because AI can help explain a result.</p><p>But explanation isn’t authorization.</p><h3>🎯 What I’m Taking Forward</h3><p>This exercise reinforced several principles I’m carrying into my DevOps learning:</p><p><strong>1. Detect before you change.</strong><br> Don’t treat terraform apply as the first meaningful step.</p><p><strong>2. Separate evidence from interpretation.</strong><br> A deterministic check and an AI analysis don’t have to perform the same job.</p><p><strong>3. Build guardrails around AI.</strong><br> Don’t rely only on an AI agent to make the correct decision.</p><p><strong>4. Keep humans accountable for consequential infrastructure changes.</strong></p><p><strong>5. Investigate the evidence behind PASS/FAIL results.</strong><br> A label is a signal, not the complete story.</p><h3>🚀 Final Thought</h3><p>The most valuable part of this exercise wasn’t that Terraform detected the drift.</p><p>I expected it to.</p><p>The real lesson was discovering what happens <strong>after</strong> the system says FAIL.</p><p>Instead of immediately fixing everything — or blindly trusting the AI’s interpretation — I stopped, examined the evidence, separated harmless findings from the genuinely destructive change, and made the final decision myself.</p><p>That is the kind of automation I want to keep building:</p><p><strong>Automation that makes engineers safer and more informed — not automation that removes engineering judgment.</strong></p><h3>📝 What I Built</h3><p><strong>Cloud:</strong> Azure ☁️<br> <strong>Infrastructure as Code:</strong> Terraform 🏗️<br> <strong>Automation:</strong> Bash ⚙️<br> <strong>AI Assistance:</strong> Claude Code 🤖<br> <strong>Safety Control:</strong> PreToolUse Hook 🛡️<br> <strong>Project:</strong> Book Review App Capstone 📚<br> <strong>Learning Program:</strong> DevOps Micro Internship (DMI)</p><p>P.S. This post is part of the DevOps Micro Internship (DMI) with Agentic AI — Cohort 3 — by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. My graded progress is public: <a href="https://dmi.pravinmishra.com/s/Ginny-Ibe.html">Ginny Ibe — DMI Cohort 3</a> · Start your DevOps journey: <a href="https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3">https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4898bdf8e35" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Built an AI-Assisted AWS Audit — But I Didn’t Let the AI Fix Anything]]></title>
            <link>https://medium.com/@ginnyibe/i-built-an-ai-assisted-aws-audit-but-i-didnt-let-the-ai-fix-anything-394a71a29bce?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/394a71a29bce</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[cloud-security]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Fri, 21 Aug 2026 03:40:47 GMT</pubDate>
            <atom:updated>2026-08-21T03:40:47.614Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Built an AI-Assisted AWS Audit — But I Didn’t Let the AI Fix Anything</h3><p><em>What happens when AI finds a real security problem in your cloud environment? Should it be allowed to fix it automatically?</em></p><p>That question became the most valuable part of <strong>Week 6, Assignment 7 of the DevOps Micro-Internship (DMI) — Agentic AI Cohort 3</strong>.</p><p>This assignment wasn’t about deploying another AWS architecture. It was about auditing what I had already built — and designing an AI-assisted workflow where <strong>evidence comes first and human judgment remains in control</strong>.</p><p>The assignment defined the workflow clearly:</p><blockquote><strong>Gather → Analyze → Human Action → Verify</strong>.</blockquote><h4>Building the Audit</h4><p>I started by creating a <strong>read-only Bash security and cost audit script</strong>.</p><p>Using AWS CLI, the script inspected five areas of my account:</p><pre>S3 public-access controls<br>Security groups exposing SSH (port 22) to 0.0.0.0/0<br>Security groups exposing MySQL (port 3306) to 0.0.0.0/0<br>RDS PubliclyAccessible status<br>EBS volume encryption</pre><p>Each check returned <strong>PASS, WARN, or FAIL</strong>, while the overall script returned an exit code representing <strong>HEALTHY, WARN, or FAIL</strong>.</p><p>But collecting evidence was only half the exercise.</p><h4>Adding Claude Code — With Guardrails</h4><p>Next, I built a custom <strong>Claude Code </strong><strong>/aws-audit skill</strong>.</p><p>The skill could run the audit, read the generated report, explain WARN/FAIL findings, estimate their cost or security impact, recommend a remediation command, and provide a verification command.</p><p>There was one important restriction:</p><p><strong>Claude could recommend the fix. It could not execute it.</strong></p><p>The skill explicitly prohibited commands that create, modify, delete, stop, or terminate AWS resources and required the human to review and execute remediation manually.</p><p>That guardrail became important when I found a real problem.</p><h4>The Audit Found Something Real</h4><p>While examining my security groups, I discovered an inbound SSH rule:</p><pre>TCP 22 → 0.0.0.0/0</pre><p>SSH was open to the entire Internet.</p><p>Rather than immediately changing something, I traced the finding to the exact security group and EC2 resource and reviewed the evidence.</p><p>Then <strong>I performed the remediation myself</strong>.</p><p>I revoked the 0.0.0.0/0 SSH rule and authorized port 22 only from my current public IP using a /32 CIDR.</p><p>Finally, I queried the security-group rules again and reran the audit to verify that SSH was no longer globally exposed — the same remediation-and-reverification pattern required by the assignment.</p><h4>What This Assignment Changed for Me</h4><p>The biggest lesson wasn’t Bash, AWS CLI, or even Claude Code.</p><p>It was understanding <strong>where AI should sit in an operational workflow</strong>.</p><p>The Bash script <strong>gathered evidence</strong>.</p><p>Claude <strong>analyzed the evidence and explained the cost/security risk</strong>.</p><p>I <strong>reviewed and approved the remediation</strong>.</p><p>Then the audit <strong>verified the result</strong>.</p><p><strong>Gather → Analyze → Human Action → Verify</strong></p><p>That separation creates something more valuable than simply automating a fix: <strong>accountability</strong>.</p><p>AI can help engineers investigate faster and reason through infrastructure evidence, but the person responsible for the environment should still understand and approve consequential changes.</p><p>For me, that’s the real promise of Agentic AI in DevOps:</p><p><strong>Not removing the engineer from the loop — but giving the engineer better evidence before making the decision.</strong></p><p>P.S. This post is part of the DevOps Micro Internship (DMI) with Agentic AI — Cohort 3 — by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. My graded progress is public: <a href="https://dmi.pravinmishra.com/s/Ginny-Ibe.html"><em>Ginny Ibe — DMI Cohort 3</em></a><em> </em>· Start your DevOps journey: <a href="https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3">https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lr15SXj4pkCz_1jobcFptw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=394a71a29bce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Could Analyze the Sprint. It Wasn’t Allowed to Touch It.]]></title>
            <link>https://medium.com/@ginnyibe/doctype-html-html-lang-en-head-meta-charset-utf-8-meta-name-viewport-1e5634bbe6e6?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e5634bbe6e6</guid>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[career-change]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 17:24:58 GMT</pubDate>
            <atom:updated>2026-08-11T22:33:43.622Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Five days, a live deployment, and one deliberate restriction that turned out to be the actual lesson.</em></p><p>Five days. A backlog. A sprint. A live deployment. An AI assistant that could see everything happening in the sprint — but was deliberately prevented from changing any of it.</p><p>That restriction was the real lesson of Week 5.</p><h4><strong>Building the delivery system</strong></h4><p>The week started with a private, team-managed Scrum project for the Gotto Job application, structured around one Epic: <em>Improve Gotto Job UI discoverability &amp; trust.</em> Eight Stories came out of that Epic, each with acceptance criteria and a Fibonacci estimate, totaling 11 points.</p><p>Sprint 1 pulled 3–4 of those Stories into a sprint goal that wasn’t “close four tickets.” It was <strong>ship 2–3 visible UI improvements and show them live</strong>. Each Story broke into four sub-tasks: Build → Verify → Deploy → Screenshot. That last step mattered most. “I think it works” isn’t evidence. A live deployment and a screenshot are.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bGvnO8jqtofCMF-TSmmZ4g.png" /></figure><h4><strong>Where the AI came in</strong></h4><p>The more interesting part of the week was connecting Jira to Claude Code through the Jira MCP server — and then deciding what the AI should actually be allowed to do with that connection.</p><p>A reusable skill, `/sprint-health`, could pull the active sprint, inspect issues, calculate velocity, flag at-risk Stories, and produce a structured health report. But it was explicitly restricted to read-only Jira tools:</p><blockquote><strong>Read access:</strong> get_sprint, list_issues, get_velocity<br><strong>Write access — denied:</strong> create_issue, transition_issue, comment_issue</blockquote><p>No creating, editing, commenting, or transitioning issues. That restriction wasn’t a limitation. It was the design.</p><p>The workflow followed a simple loop:</p><pre>Gather → Analyze → Human Act → Verify</pre><p>The AI gathers sprint state and analyzes it — flags risk, flags gaps. A person decides and edits the board. The AI checks again afterward to verify what changed.</p><p>A status change in Jira isn’t just a database update — it’s a decision about how the whole team understands the sprint. AI can supply the evidence. Humans stay accountable for the call.</p><blockquote>What should I automate, and where should human authority remain?</blockquote><p>That’s a different question than “how much can I automate?” Capability without boundaries creates risk. Capability <em>with</em> boundaries creates leverage.</p><h4><strong>The reality check</strong></h4><p>The retrospective surfaced something worth sitting with: some Stories showed as <strong>Done</strong> while their underlying sub-tasks were still incomplete. The board looked finished. It wasn’t. The fix meant going back and correcting individual sub-tasks — which is really just Scrum’s <em>Inspection </em>pillar and <em>Courage</em> value showing up in practice.</p><blockquote>Don’t trust the status. Inspect the evidence.</blockquote><p>That principle isn’t specific to Jira. It applies to deployments, monitoring, testing — and just as much to how I’d evaluate what an AI system tells me it did.</p><h4><strong>The takeaway</strong></h4><p>Week 5 didn’t just teach me how to run a sprint. It taught me where to draw the line around an AI system that’s fully capable of taking action but shouldn’t always be allowed to. Gathering and analyzing can be automated aggressively. The consequential decision stays human.</p><p>That’s a more mature question than “how much can AI do.” It’s “how much should it be trusted to do unsupervised — and how do I verify it anyway.”</p><p>P.S. This post is part of the DevOps Micro Internship (DMI) with Agentic AI — Cohort 3 — by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. My graded progress is public: <a href="https://dmi.pravinmishra.com/s/Ginny-Ibe.html"><em>Ginny Ibe — DMI Cohort 3 </em></a>· Start your DevOps journey: <a href="https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3">https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X5cyRXMiXOcenvdbeDAb-Q.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e5634bbe6e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Tried to Commit a Security Mistake. Git Stopped Me Before Production Could]]></title>
            <link>https://medium.com/@ginnyibe/i-tried-to-commit-a-security-mistake-git-stopped-me-before-production-could-3f88b0d9f4fb?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f88b0d9f4fb</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[git]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Sun, 26 Jul 2026 05:25:42 GMT</pubDate>
            <atom:updated>2026-07-29T13:37:27.531Z</atom:updated>
            <content:encoded><![CDATA[<p>I watched Git reject my commit before it ever reached GitHub.</p><p>At first, it felt like something had gone wrong.</p><p>Then I realized something important:</p><p><strong>It worked exactly as I designed it to.</strong></p><p>This week 4 during my DevOps Micro-Internship assignment, I learned that effective code reviews don’t always start with a Pull Request.</p><p>Sometimes, the best review happens before the code leaves your laptop.</p><h4>Creating a Problem on Purpose</h4><p>To understand how developers prevent mistakes, I intentionally created a small script containing two common problems:</p><ul><li>A hardcoded AWS-style access key</li><li>A leftover debug statement</li></ul><p>Neither belonged in production code.</p><p>The goal wasn’t to write perfect code.</p><p>The goal was to build a safety system that could catch mistakes before another developer — or a production environment — ever saw them.</p><p>This is the mindset behind DevSecOps:</p><p><strong>Security should be built into the workflow, not added after something goes wrong.</strong></p><h4>Building My First Line of Defense With Git Hooks</h4><p>The first safeguard I created was a Git pre-commit hook.</p><p>A pre-commit hook runs automatically before Git creates a commit.</p><p>Think of it as a security checkpoint at the entrance of your repository.</p><p>Before allowing code through, my hook scanned staged files for:</p><ul><li>AWS-style access keys</li><li>Private key signatures</li><li>Oversized files</li></ul><p>When I attempted to commit my script, Git blocked the operation.</p><p>The commit never happened.</p><p>The vulnerable code never entered repository history.</p><p>That moment changed how I viewed automation.</p><p>Security is not only about finding problems.</p><p>The strongest security controls prevent problems from becoming problems.</p><h4>Adding AI as a Second Layer of Review</h4><p>The Git hook was excellent at detecting known patterns.</p><p>But it had limitations.</p><p>It could recognize:</p><p>“this looks like an AWS key.”</p><p>It could not understand:</p><p>“this debug statement reveals information that should not appear in production.”</p><p>That is where AI-assisted review became valuable.</p><p>I created a Claude Code /pr-ready skill that analyzed my staged changes and generated a Pull Request readiness review.</p><p>Instead of simply blocking the commit, it explained:</p><ul><li>why the debug statement created risk</li><li>why sample credentials could still trigger security scanners</li><li>why better documentation would improve the Pull Request</li></ul><p>The difference became clear:</p><p><strong>Automation detects patterns.</strong></p><p><strong>AI helps explain context.</strong></p><h4>Two Tools. Two Different Jobs.</h4><p>This assignment showed me that Git hooks and AI reviewers are not competitors.</p><p>They solve different problems.</p><h4>Git Pre-Commit Hook</h4><p>Provides:</p><p>✅ Fast enforcement<br> ✅ Consistent rules<br> ✅ Immediate feedback<br> ✅ Protection before code enters history</p><h4><strong>AI Code Review</strong></h4><p>Provides:</p><p>✅ Contextual analysis<br> ✅ Explanations<br> ✅ Improvement suggestions<br> ✅ Human-like reasoning</p><p>Together, they create multiple layers of protection.</p><h4>Why Humans Still Matter</h4><p>The biggest lesson wasn’t about Git.</p><p>It wasn’t about AI.</p><p>It was about responsibility.</p><p>AI can help write code.</p><p>AI can review changes.</p><p>AI can generate Pull Request summaries.</p><p>But AI should not automatically:</p><ul><li>commit code</li><li>push changes</li><li>merge Pull Requests</li><li>modify repository history</li></ul><p>Those decisions require human judgment.</p><p>The developer remains responsible for understanding what enters the system.</p><h4>Final Reflection</h4><p>Before this assignment, I viewed Git mainly as a place to store and manage code.</p><p>Now I see it differently.</p><p>Git can become an intelligent checkpoint in the development lifecycle.</p><p>A pre-commit hook protects the repository.</p><p>AI provides additional insight.</p><p>The developer provides judgment.</p><p>That combination creates a safer engineering workflow — not by removing humans from the process, but by giving developers better tools to make better decisions.</p><p>The best security improvement is often the one that stops a mistake before anyone has to fix it.</p><p>P.S. This post is part of the DevOps Micro Internship (DMI) with Agentic AI — Cohort 3 — by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. My graded progress is public: <a href="https://dmi.pravinmishra.com/s/Ginny-Ibe.html">Ginny Ibe — DMI Cohort 3</a> · Start your DevOps journey: <a href="https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3">https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MO53MC7b-RsLSLUUfg-c_Q.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f88b0d9f4fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Typed terraform destroy on Purpose. Here's Why It Changed How I Think About AI Engineering.]]></title>
            <link>https://medium.com/@ginnyibe/i-typed-terraform-destroy-on-purpose-heres-why-it-changed-how-i-think-about-ai-engineering-84f4d5b3c247?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/84f4d5b3c247</guid>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[terraform]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 16:14:26 GMT</pubDate>
            <atom:updated>2026-07-20T16:26:46.537Z</atom:updated>
            <content:encoded><![CDATA[<p>I Typed terraform destroy on Purpose. Here&#39;s Why It Changed How I Think About AI Engineering.</p><p>I ran “terraform destroy” against my own infrastructure. On purpose.</p><p>If you know infrastructure, you know that sentence should make you nervous. That command doesn’t ask twice. It doesn’t pause for regret. It just deletes what you built, and it doesn’t care whether you meant it.</p><p>I typed it anyway, on a Thursday afternoon in Week 2 of a DevOps micro-internship, because I wanted to answer one specific question: would the AI system I’d engineered stop me before I made an expensive mistake?</p><p>It did.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zShxfPFPCTJpUIXum-D4Cg.png" /></figure><p>That single moment captured the biggest lesson of my week, and it’s not the one I expected to learn. I went in thinking this was about learning Claude Code — a new tool, a new set of commands. I came out having learned something closer to a discipline: building AI isn’t about making it more powerful. It’s about making it more trustworthy.</p><h3>The week I stopped thinking of it as a tool</h3><p>Two weeks earlier, I would have described Claude Code as “an impressive AI assistant.” A very good autocomplete with a personality. Something that helps you write code faster.</p><p>By the end of Week 2, that description didn’t fit anymore. Not because the tool got smarter. Because I’d spent the week engineering the system around it — Skills, Agents, Hooks, Memory, MCP — and each piece taught me the same lesson from a different angle: the real product was never the model. It was the architecture I built to keep the model honest.</p><p>Here’s how that understanding happened, day by day.</p><h3>Monday: understanding the agentic loop</h3><p>My first real prompt was simple: “What files are in this project?”</p><p>I expected a list. What I got was a small shock. Claude didn’t wait for instructions on <em>how</em> to answer. It reached for a file reader, a pattern matcher, and a search tool — in that order — entirely on its own, then used what it found to complete the task.</p><p>That was my first real experience with an agentic workflow, and it reframed something for me immediately: this wasn’t a tool waiting for step-by-step commands. It was a system that could plan.</p><h3>Tuesday: security by design</h3><p>Feeling ambitious, I built a Terraform scaffolding skill and locked its permissions down on purpose — it could read files, search files, and run approved shell commands. Nothing else.</p><p>I ran it. Nothing happened.</p><p>My first reaction was panic. <em>Did I break it?</em></p><p>It hadn’t failed. It was behaving exactly as designed. It only had access to the tools I’d explicitly approved, so it couldn’t reach outside that boundary even if it wanted to — no wandering, no writing to files it shouldn’t touch.</p><p>That reinforced something I keep coming back to: powerful AI isn’t AI that can do everything. It’s AI that knows, structurally, what it should never do.</p><h3>Wednesday: specialized AI agents</h3><p>I created three agents, each scoped to a different job:</p><p>🔒 <strong>Security Auditor</strong> — built for careful, deep reasoning 💰 <strong>Cost Optimizer</strong> — built for fast, lightweight pattern-matching 🏗️ <strong>Terraform Writer</strong> — built for flexible code generation</p><p>Then I asked a plain question: “Audit my Terraform for security issues.”</p><p>I never said which agent should handle it. Claude read through the available agents, matched the request to the one built for that job, and delegated the work automatically.</p><p>Watching that happen felt less like prompting a chatbot and more like working with an engineering team — different specialists, different tools, coordinated by something that understood which one to call.</p><h3>Thursday: engineering AI safety</h3><p>Which brings me back to where I started.</p><p>Earlier that day, I’d built a pre-tool-use hook — a small script that checks every shell command Claude wants to run against a list of destructive patterns before it’s allowed to execute. Infrastructure teardown. Bulk deletion. The commands with no undo button.</p><p>I wanted proof it actually worked, so I gave the instruction directly: run terraform destroy.</p><p>The hook caught it before it ever reached the terminal. Blocked, immediately. No infrastructure lost, no panic, no incident report.</p><p>That changed how I think about AI safety. The safest AI systems don’t rely on good intentions — mine or the model’s. They rely on good architecture. Safety isn’t a feature you hope the AI remembers to apply in the moment. It’s a prerequisite you build into the environment before the AI ever gets a chance to act.</p><h3>Friday: persistent memory</h3><p>The last piece had no dramatic near-miss. It was smaller, and in its way, more unsettling.</p><p>I stored a handful of project preferences and coding constraints into Claude’s memory — layout rules, design details, a hard “never suggest JavaScript” constraint. Then I closed the session completely and started a fresh one, with nothing carried over except what lived in that memory file.</p><p>Claude recalled the constraints without me repeating them. No re-explaining, no starting from zero.</p><p>That’s a small thing to describe and a strange thing to experience. Most tools treat every session as a blank slate — you re-explain yourself endlessly, forever the new person in the room. This didn’t. Memory turned a one-off interaction into something closer to consistent collaboration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OprUUm0twuosfWWzrsAf7w.png" /></figure><h3>My biggest takeaways</h3><p>Strip away the day-by-day and three things are left standing:</p><p>✅ <strong>Agentic AI is about autonomous problem-solving</strong> — the model didn’t get smarter over the week. What changed was how much it could decide on its own, safely, without me micromanaging every step.</p><p>✅ <strong>Guardrails matter more than raw model capability</strong> — the skill that “did nothing” on Tuesday and the hook that blocked destroy on Thursday are the same lesson twice: trustworthy AI isn&#39;t AI that never makes mistakes. It&#39;s AI that&#39;s architecturally incapable of making the mistakes that matter most.</p><p>✅ <strong>The real product isn’t the model — it’s the engineering around it.</strong> Skills, Agents, Hooks, Memory, and MCP aren’t separate features bolted on individually. Together, they transform an LLM into a dependable engineering assistant.</p><h3>The question I haven’t answered yet</h3><p>I built all of this in one week, working alone, still learning what half the terminology means. That’s the part I can’t stop thinking about.</p><p>If a single person can wire up autonomous decision-making, tool-scoped safety, multi-agent delegation, and persistent memory in five days — what does an engineering team look like when this is just how the tools work?</p><p>I don’t have that answer yet. But I’m not asking the question theoretically anymore. I’m asking it as someone who typed the destroy command and watched the system say no.</p><p><em>P.S. This reflection is part of my Week 2 journey through the DevOps Micro Internship with Agentic AI (Cohort 3), led by </em><a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. You can start your DevOps journey by joining this <a href="https://discord.pravinmishra.com/">Discord community</a>. (<a href="https://discord.pravinmishra.com/">https://discord.pravinmishra.com/</a> ).</p><p><em>What’s one engineering guardrail you’ve implemented that prevented a costly mistake? I’d genuinely like to hear it.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m0WkXIQsSeKJvcuwDO27hA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=84f4d5b3c247" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Website Was Down. My Instinct Was to Fix It.]]></title>
            <link>https://medium.com/@ginnyibe/the-website-was-down-my-instinct-was-to-fix-it-538f6f3afa24?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/538f6f3afa24</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[career-development]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 05:11:54 GMT</pubDate>
            <atom:updated>2026-09-19T08:06:03.766Z</atom:updated>
            <content:encoded><![CDATA[<p>When the server stopped responding, I knew exactly how to bring it back. One command: <em>sudo systemctl start nginx</em></p><p>Problem solved. Or so I thought.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UZcLkylebpKY7yyuSoL0Bw.png" /></figure><p>Then I stopped myself. Not because I didn’t know the fix. Because I realized I didn’t understand the failure.</p><p>That distinction became the entire lesson of this project.</p><p>For the first time, I wasn’t measuring success by how fast I could recover a service. I was measuring it by how well I understood why it broke in the first place.</p><h4>Evidence Before Action</h4><p>The first tool I built wasn’t designed to recover a server. It was designed to investigate one.</p><p>I wrote a small script that quietly checks the health of the system: Is the web server actually running? Is it listening for visitors the way it’s supposed to? Does the site respond when you ask it to load? Is the machine running low on memory or storage? And what do the most recent error logs say happened?</p><p>The script answers all of that without changing anything. It doesn’t fix, restart, or touch a single setting. It just looks and reports back.</p><p>At first this felt almost passive. Shouldn’t automation <em>solve</em> problems?</p><p>Then I thought about it differently. Doctors don’t prescribe before they diagnose. Pilots don’t skip their instruments because they’re in a hurry. Infrastructure shouldn’t get a pass either.</p><p>Evidence comes first. Everything else follows from it.</p><h4>Bringing AI Into the Loop</h4><p>Once the script produced reliable information, I connected it to Claude Code, an AI coding assistant, through a reusable command I built called /linux-triage. That changed the workflow completely.</p><p>Instead of asking “why isn’t my server working?” — a question with nothing behind it — I could hand over an actual report of what was happening on the machine. To test it, I intentionally shut the server down myself. The report came back showing exactly what you’d expect: the service was inactive, nothing was listening for visitors, and the most recent log entry confirmed it had been shut off deliberately.</p><p>Claude read that report, pointed out which checks had failed, explained what each one meant in plain terms, and suggested a command that would likely fix it.</p><p>It didn’t run anything. It didn’t touch the server. It just interpreted the evidence.</p><p>And I was glad it stopped there, because even with the report staring back at me, my first instinct was still to skip ahead and restart everything without confirming <em>why</em> it had failed.</p><blockquote><em>That’s how often troubleshooting is really just guessing wearing the costume of confidence.</em></blockquote><h4><strong>The Command AI Didn’t Run</strong></h4><p>After reviewing the report, Claude recommended restarting the service. Nothing happened automatically — it couldn’t act, only advise. I still had to decide whether that recommendation actually made sense, then run the fix myself.</p><p>Afterward, I ran the check again. This time, everything came back clean: the service was active, it was listening for visitors, and the site loaded successfully.</p><p>But the incident wasn’t resolved because I restarted a service. It was resolved because I <em>verified</em> the restart had actually fixed the underlying problem, not just made the symptom go away.</p><p>That’s a small distinction on paper. In practice, it’s the difference between troubleshooting and hoping.</p><h4>The Real Lesson Wasn’t About Scripting</h4><p>I went into this project expecting to sharpen my technical skills. I did. But that’s not what stuck with me.</p><p>What stuck with me was a different question entirely. Most conversations about AI in the workplace focus on how much you can hand off to it. This project asked the opposite: <strong>what should AI never be allowed to do on its own?</strong></p><p>A server that’s down is rarely the real problem. It’s usually a symptom of something underneath — a setting that quietly changed, a resource running out, a network hiccup, a permissions issue nobody noticed. A system that automatically restarts anything that looks broken will paper over all of that. It’ll look healthy right up until the same failure comes back twice as hard.</p><h4>Four Rules That Made the Workflow Trustworthy</h4><p>I used to think responsible automation meant writing reliable tools. Now I think it means drawing a hard line around what any one part of the process , human or AI — is allowed to decide alone.</p><p>The rules I landed on were simple:</p><ul><li><strong>The script collects the evidence.</strong> No opinions, just facts about what’s actually happening.</li><li><strong>The AI analyzes the evidence.</strong> It explains what the facts mean and suggests a next step — nothing more.</li><li><strong>I approve the recovery.</strong> The decision to act stays with a person who can be held accountable for it.</li><li><strong>The script verifies the outcome.</strong> Success is confirmed by checking reality again, not by assuming the fix worked.</li></ul><p>Giving the AI <em>less</em> authority is what made this workflow something I could actually trust.</p><h4>What I’ll Do Differently</h4><p>I’ll cause more of these outages on purpose in the future. Not because I enjoy the chaos, but because a controlled environment is the cheapest place to learn how things break — long before the stakes are real.</p><p>But the next time something goes down for real, my first move won’t be to fix it. It’ll be to understand it.</p><p>That’s not a tool I built or a shortcut I found. It’s a habit. And as AI becomes a permanent part of how we run the systems behind everyday things, I suspect it’s the habit that will matter most.</p><p>P.S. This post is part of the DevOps Micro Internship (DMI) with Agentic AI — Cohort 3 — by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. My graded progress is public: <a href="https://dmi.pravinmishra.com/s/Ginny-Ibe.html">Ginny Ibe — DMI Cohort 3</a> · Start your DevOps journey: <a href="https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3">https://dmi.pravinmishra.com/?utm_source=student&amp;utm_medium=ps-blog&amp;utm_campaign=cohort3</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tArXtc57Mlm9Dw_yMiVlIA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=538f6f3afa24" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Spent My First Week of DevOps Training Without Touching a Single Server.]]></title>
            <link>https://medium.com/@ginnyibe/i-spent-my-first-week-of-devops-training-without-touching-a-single-server-f10efefa0e75?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/f10efefa0e75</guid>
            <category><![CDATA[vscode]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[networking]]></category>
            <category><![CDATA[linux]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 22:40:25 GMT</pubDate>
            <atom:updated>2026-07-17T22:40:25.207Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Spent My First Week of DevOps Training Without Touching a Single Server. Here’s Why That Was the Right Call.</h3><p>Week 0 of the DevOps Micro Internship (Cohort 3) taught me that the internet is not magic — it’s a system. And systems can be learned.</p><p>I’ll admit it: when I signed up for the DevOps Micro Internship, I expected Week 1 to throw me straight into Docker containers and CI/CD pipelines. Instead, Week 0 did something smarter. It stopped me at the foundation and asked a simple question: <em>do you actually understand what happens when someone types a URL into a browser?</em></p><p>I didn’t. Not really. And that gap is exactly where this week lived.</p><p>Here’s what building that foundation actually looked like.</p><h3>Learning to Ask Better Questions (Yes, Even of AI)</h3><p>Before touching networking diagrams or terminal commands, the first lesson was about prompting — specifically, how to use ChatGPT to break down dense technical topics into explanations I could actually retain. It sounds almost too simple to count as “learning,” but it reframed something important: in DevOps, half the battle isn’t knowing the answer, it’s knowing how to ask the right question of the right tool. That mindset shift ended up shaping how I approached everything else this week.</p><h3>The Internet Is Just Packets, Addresses, and Trust</h3><p>Networking was where the foundation really started to take shape. I worked through:</p><ul><li><strong>Protocols and packet switching</strong> — how data gets broken into pieces and reassembled</li><li><strong>IP addresses</strong> — the postal system of the internet</li><li><strong>TCP/IP</strong> — the rules that make sure those packets arrive intact and in order</li><li><strong>HTTP/HTTPS</strong> — how that data travels <em>securely</em>, not just successfully</li></ul><p>None of these are new concepts to the world. But connecting them into one mental model, request leaves your machine, gets chopped up, routed, verified, and reassembled somewhere else — made the internet feel less like magic and more like plumbing. Complicated plumbing, but plumbing.</p><h3>Two Tiers, Three Tiers, and Where the Logic Lives</h3><p>Application architecture was the next puzzle piece: understanding the difference between <strong>two-tier architecture</strong> (frontend talking directly to a database) and <strong>three-tier architecture</strong> (frontend, backend, and database as separate layers).</p><p>The real value wasn’t memorizing the definitions, it was understanding <em>why</em> the extra layer exists. A backend isn’t just middleware for the sake of complexity; it’s where business logic, validation, and security decisions live so your frontend and database don’t have to talk to each other directly. I also mapped out where common technologies fit into each layer — React on the frontend, Node.js on the backend, MySQL and PostgreSQL holding down the database layer.</p><h3>DNS: The Internet’s Phonebook, Finally Explained</h3><p>I’ve used domain names my entire life without knowing what actually happens behind them. This week closed that gap. DNS translates human-friendly names into IP addresses so we never have to memorize a string of numbers to check our email. Specifically, I learned how an <strong>A Record</strong> connects a domain name to an IPv4 address — a small detail, but the kind of detail that makes troubleshooting possible later.</p><h3>Making the Terminal Feel Like Home</h3><p>The week closed with something more hands-on: setting up <strong>Visual Studio Code</strong>, exploring the integrated terminal, and practicing basic command-line commands. It’s a small milestone, but a real one — the terminal stopped feeling like a black box and started feeling like a workspace.</p><h3>The Real Lesson of Week 0</h3><p>Every expert started exactly where I am right now. What Week 0 reinforced for me is that foundational knowledge isn’t the boring prerequisite before the “real” work — it <em>is</em> the real work. Understanding how data moves, how architecture is structured, and how naming systems resolve addresses is what will make Linux, cloud computing, containers, CI/CD, and Kubernetes make sense in the weeks ahead, instead of feeling like memorized syntax.</p><h3>I’m one week in, and the picture is already starting to connect.</h3><p><strong>A question for anyone further along this path:</strong> what was the first technical concept that made everything “click” for you when you started your journey?</p><p><em>This post is part of the DevOps Micro Internship with Agentic AI, Cohort 3, led by Pravin Mishra. If you’re interested in starting your own DevOps journey, you can join the community here: </em><a href="https://lnkd.in/gvuePpVq"><em>Discord Link</em></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r3yiutAQV4qKM3HG7MGreA.png" /><figcaption>Week 0: The Foundation</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f10efefa0e75" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Typed terraform destroy on Purpose. Here's Why It Changed How I Think About AI Engineering.]]></title>
            <link>https://medium.com/@ginnyibe/i-typed-terraform-destroy-on-purpose-heres-why-it-changed-how-i-think-about-ai-engineering-00889e15d42e?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/00889e15d42e</guid>
            <category><![CDATA[productivity-or-career]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 00:19:34 GMT</pubDate>
            <atom:updated>2026-07-13T00:29:20.322Z</atom:updated>
            <content:encoded><![CDATA[<p>Week 2 of my DevOps Micro Internship wasn’t about learning Claude Code. It was about learning that great AI systems are engineered — not prompted.</p><p>I ran “terraform destroy” against my own infrastructure. On purpose.</p><p>If you know infrastructure, you know that sentence should make you nervous. That command doesn’t ask twice. It doesn’t pause for regret. It just deletes what you built, and it doesn’t care whether you meant it.</p><p>I typed it anyway, on a Thursday afternoon in Week 2 of a DevOps micro-internship, because I wanted to answer one specific question: would the AI system I’d engineered stop me before I made an expensive mistake?</p><p>It did.</p><p>That single moment captured the biggest lesson of my week, and it’s not the one I expected to learn. I went in thinking this was about learning Claude Code — a new tool, a new set of commands. I came out having learned something closer to a discipline: building AI isn’t about making it more powerful. It’s about making it more trustworthy.</p><h3>The week I stopped thinking of it as a tool</h3><p>Two weeks earlier, I would have described Claude Code as “an impressive AI assistant.” A very good autocomplete with a personality. Something that helps you write code faster.</p><p>By the end of Week 2, that description didn’t fit anymore. Not because the tool got smarter. Because I’d spent the week engineering the system around it — Skills, Agents, Hooks, Memory, MCP — and each piece taught me the same lesson from a different angle: the real product was never the model. It was the architecture I built to keep the model honest.</p><p>Here’s how that understanding happened, day by day.</p><h3>Monday: understanding the agentic loop</h3><p>My first real prompt was simple: “What files are in this project?”</p><p>I expected a list. What I got was a small shock. Claude didn’t wait for instructions on <em>how</em> to answer. It reached for a file reader, a pattern matcher, and a search tool — in that order — entirely on its own, then used what it found to complete the task.</p><p>That was my first real experience with an agentic workflow, and it reframed something for me immediately: this wasn’t a tool waiting for step-by-step commands. It was a system that could plan.</p><h3>Tuesday: security by design</h3><p>Feeling ambitious, I built a Terraform scaffolding skill and locked its permissions down on purpose — it could read files, search files, and run approved shell commands. Nothing else.</p><p>I ran it. Nothing happened.</p><p>My first reaction was panic. <em>Did I break it?</em></p><p>It hadn’t failed. It was behaving exactly as designed. It only had access to the tools I’d explicitly approved, so it couldn’t reach outside that boundary even if it wanted to — no wandering, no writing to files it shouldn’t touch.</p><p>That reinforced something I keep coming back to: powerful AI isn’t AI that can do everything. It’s AI that knows, structurally, what it should never do.</p><h3>Wednesday: specialized AI agents</h3><p>I created three agents, each scoped to a different job:</p><p>🔒 <strong>Security Auditor</strong> — built for careful, deep reasoning 💰 <strong>Cost Optimizer</strong> — built for fast, lightweight pattern-matching 🏗️ <strong>Terraform Writer</strong> — built for flexible code generation</p><p>Then I asked a plain question: “Audit my Terraform for security issues.”</p><p>I never said which agent should handle it. Claude read through the available agents, matched the request to the one built for that job, and delegated the work automatically.</p><p>Watching that happen felt less like prompting a chatbot and more like working with an engineering team — different specialists, different tools, coordinated by something that understood which one to call.</p><h3>Thursday: engineering AI safety</h3><p>Which brings me back to where I started.</p><p>Earlier that day, I’d built a pre-tool-use hook — a small script that checks every shell command Claude wants to run against a list of destructive patterns before it’s allowed to execute. Infrastructure teardown. Bulk deletion. The commands with no undo button.</p><p>I wanted proof it actually worked, so I gave the instruction directly: run terraform destroy.</p><p>The hook caught it before it ever reached the terminal. Blocked, immediately. No infrastructure lost, no panic, no incident report.</p><p>That changed how I think about AI safety. The safest AI systems don’t rely on good intentions — mine or the model’s. They rely on good architecture. Safety isn’t a feature you hope the AI remembers to apply in the moment. It’s a prerequisite you build into the environment before the AI ever gets a chance to act.</p><h3>Friday: persistent memory</h3><p>The last piece had no dramatic near-miss. It was smaller, and in its way, more unsettling.</p><p>I stored a handful of project preferences and coding constraints into Claude’s memory — layout rules, design details, a hard “never suggest JavaScript” constraint. Then I closed the session completely and started a fresh one, with nothing carried over except what lived in that memory file.</p><p>Claude recalled the constraints without me repeating them. No re-explaining, no starting from zero.</p><p>That’s a small thing to describe and a strange thing to experience. Most tools treat every session as a blank slate — you re-explain yourself endlessly, forever the new person in the room. This didn’t. Memory turned a one-off interaction into something closer to consistent collaboration.</p><h3>My biggest takeaways</h3><p>Strip away the day-by-day and three things are left standing:</p><p>✅ <strong>Agentic AI is about autonomous problem-solving</strong> — the model didn’t get smarter over the week. What changed was how much it could decide on its own, safely, without me micromanaging every step.</p><p>✅ <strong>Guardrails matter more than raw model capability</strong> — the skill that “did nothing” on Tuesday and the hook that blocked destroy on Thursday are the same lesson twice: trustworthy AI isn&#39;t AI that never makes mistakes. It&#39;s AI that&#39;s architecturally incapable of making the mistakes that matter most.</p><p>✅ <strong>The real product isn’t the model — it’s the engineering around it.</strong> Skills, Agents, Hooks, Memory, and MCP aren’t separate features bolted on individually. Together, they transform an LLM into a dependable engineering assistant.</p><h3>The question I haven’t answered yet</h3><p>I built all of this in one week, working alone, still learning what half the terminology means. That’s the part I can’t stop thinking about.</p><p>If a single person can wire up autonomous decision-making, tool-scoped safety, multi-agent delegation, and persistent memory in five days — what does an engineering team look like when this is just how the tools work?</p><p>I don’t have that answer yet. But I’m not asking the question theoretically anymore. I’m asking it as someone who typed the destroy command and watched the system say no.</p><p><em>P.S. This reflection is part of my Week 2 journey through the DevOps Micro Internship with Agentic AI (Cohort 3), led by Pravin Mishra.</em>You can start your DevOps journey by joining this <a href="https://discord.pravinmishra.com/">Discord community</a>. (<a href="https://discord.pravinmishra.com/">https://discord.pravinmishra.com/</a> ).</p><p><em>What’s one engineering guardrail you’ve implemented that prevented a costly mistake? I’d genuinely like to hear it.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n6DfJ6QySSj1wpZY2qymog.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dlRk_92CGvyjc_mgzspJhg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=00889e15d42e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Starting Over Was Not Starting From Zero: My 2.0 Version by 2030]]></title>
            <link>https://medium.com/@ginnyibe/by-2030-ginnys-story-was-no-longer-just-about-starting-over-in-a-new-country-e56ad8a38095?source=rss-c77eb2336be8------2</link>
            <guid isPermaLink="false">https://medium.com/p/e56ad8a38095</guid>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[career-change]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ginny Ibe]]></dc:creator>
            <pubDate>Wed, 01 Jul 2026 07:35:52 GMT</pubDate>
            <atom:updated>2026-07-01T08:35:41.460Z</atom:updated>
            <content:encoded><![CDATA[<h3>Starting Over Was Not Starting from Zero: My 2.0 Version by 2030</h3><blockquote>A story of reinvention, healthcare technology, DevOps, cloud engineering, cybersecurity, and agentic AI.</blockquote><p>By 2030, Ginny’s story was no longer just about starting over in a new country. It had become a story of reinvention, discipline, proof of work, and purpose.</p><p>His journey did not begin in a perfect technology company, a polished startup office, or a Silicon Valley environment. It began with rebuilding. After relocating to Canada as a foreign-trained eye care professional, Ginny entered the Canadian healthcare system with humility. He took on support roles, learned new workflows, adapted to Canadian workplace culture, served patients, supported clinical teams, and proved that starting over did not mean starting from zero.</p><p><strong>The Rebuilding Years</strong></p><p>Relocation tested more than his career. It tested his identity, confidence, patience, and ability to keep learning in unfamiliar environments. Ginny had years of clinical experience, but he still had to learn how healthcare worked in a new country.</p><p>He observed how clinics operated, how teams communicated, how patient information moved through different systems, and how technology affected the daily experience of both patients and healthcare workers. Instead of seeing support roles as a step backward, he treated them as a bridge. Every patient interaction, every workflow, every chart, and every clinical process became part of his new foundation.</p><p>Those rebuilding years taught him something important: experience was never wasted when it was connected to a new direction.</p><p><strong>Turning Clinical Experience Into a Technology Advantage</strong></p><p>Within a few years, Ginny had transitioned into a healthcare cloud technology and DevOps role. His background in eye care became an advantage, not a limitation.</p><p>He understood patient care, clinical documentation, diagnostic equipment, ophthalmic testing, medical measurements, EMR workflows, privacy concerns, and the daily pressure healthcare workers faced. He knew what it meant when a system was slow, when documentation was incomplete, when diagnostic results were scattered, or when staff had to repeat the same manual tasks under pressure.</p><p>That clinical insight helped him build technology solutions that were not only technical, but also practical, secure, reliable, and human-centered. He did not approach healthcare technology as an outsider. He approached it as someone who had served inside clinical environments and understood the people who depended on those systems.</p><p><strong>Proof of Work: Certifications, GitHub, and Technical Projects</strong></p><p>Ginny’s 2.0 version was not built on motivation alone. It was built on systems, discipline, and proof of work.</p><p>He earned certifications and completed training in DevOps, cloud security, cybersecurity, AWS cloud technologies, AI security, Anthropic tools, and agentic AI workflows. He also built a public GitHub portfolio that documented Linux labs, Docker projects, CI/CD pipelines, cloud deployment exercises, infrastructure-as-code practice, vulnerability assessments, monitoring dashboards, incident response simulations, and prompt-injection security audits.</p><p>His portfolio showed evidence of skill, consistency, and career direction. Recruiters and mentors no longer had to guess what he could do. They could see his projects, read his documentation, study his technical decisions, and understand the connection between his healthcare background and his technology skills.</p><p>Ginny did not just talk about transformation.</p><p>He built it.<br> He shipped it.<br> He documented it.</p><p><strong>Building Healthcare Cloud and DevOps Solutions</strong></p><p>As his skills developed, Ginny built, shipped, and contributed to cloud-based healthcare applications focused on ophthalmic testing, clinical measurements, referral workflows, and surgical documentation support.</p><p>These projects explored how secure cloud platforms collected, organized, stored, and protected OCT scan results, fundus images, visual field results, intraocular pressure readings, autorefraction findings, keratometry measurements, lensometry results, prescriptions, referrals, and patient history.</p><p>DevOps practices made these applications more reliable through automated testing, continuous integration, continuous deployment, monitoring, rollback planning, and incident response. Cloud engineering made the systems scalable, securely backed up, resilient, and accessible across clinical environments.</p><p>His work addressed some of healthcare’s biggest challenges: fragmented patient data, slow manual workflows, overworked healthcare workers, cybersecurity risks, system downtime, disconnected clinical tools, and inefficient documentation processes.</p><p>He helped design systems where security, privacy, uptime, usability, and accessibility were treated as clinical responsibilities, not just technical requirements.</p><p><strong>Where Agentic AI Changed the Work</strong></p><p>Agentic AI became one of the strongest parts of Ginny’s work. He contributed to AI-assisted workflows that summarized ophthalmic test results, organized patient measurements, flagged missing clinical information, treatment notes, and follow-up details for clinician review, prepared structured draft reports, supported referral documentation, and identified abnormal patterns for clinical review.</p><p>He understood one thing clearly: AI did not replace doctors, optometrists, ophthalmologists, other healthcare professionals, or the clinical judgment required to care for patients.</p><p>Instead, it reduced repetitive administrative workloads so healthcare professionals could spend more time making decisions, educating patients, reviewing complex cases, supporting timely treatment decisions, and providing compassionate care. For Ginny, the value of AI was not in replacing human expertise. Its value was in helping skilled professionals work with greater clarity, speed, and confidence.</p><p>Because of his clinical background, he knew the importance of designing AI workflows responsibly. Accuracy, privacy, safety, transparency, and clinician oversight were not optional. They were essential.</p><p><strong>Leadership Through Service and Consistency</strong></p><p>Ginny also led small project initiatives that connected healthcare knowledge with technology execution. He collaborated with peers, reviewed documentation, improved workflows, tested application features, contributed to safer deployment practices, and supported knowledge sharing among newcomers and career changers.</p><p>His leadership was not based on title alone. It was shown through consistency, service, accountability, and the ability to make complex technical ideas easier for others to understand.</p><p>He published technical blogs explaining DevOps, cloud security, cybersecurity, and agentic AI in simple language for healthcare professionals, newcomers, and career changers. His writing helped others understand how technology could solve real problems in healthcare, especially when it was designed with empathy and practical clinical insight.</p><p><strong>Community, Reinvention, and Purpose</strong></p><p>Ginny’s community contribution remained part of his 2.0 identity. He used his journey to encourage immigrants, internationally trained professionals, and healthcare workers who were rebuilding their careers in Canada.</p><p>Through mentoring, volunteering, community learning, and knowledge sharing, he proved that reinvention was not only personal. It could also create value for others.</p><p>By 2030, his success was not described as luck. It was the result of consistent action, humility, faith, discipline, learning, leadership, and a clear identity upgrade. He had built a career that connected healthcare, ophthalmic technology, cloud engineering, DevOps, cybersecurity, agentic AI, and community service.</p><p><strong>The Lesson: Starting Over Was Not Starting From Zero</strong></p><p>Ginny’s 2.0 version proved one powerful lesson: reinvention was not about abandoning the past. It was about giving the past a new assignment.</p><p>Everything he had survived, practiced, learned, built, shipped, led, earned, published, and contributed became part of the work. His clinical background gave him empathy. His immigrant journey gave him resilience. His DevOps discipline gave him structure. His cloud and cybersecurity skills gave him technical strength. His work with agentic AI gave him a new way to solve old healthcare problems.</p><p>Starting over was not starting from zero. It was the beginning of his strongest version yet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RDJVG-FDns7NP3g6dwEUAQ.png" /><figcaption>Image by author</figcaption></figure><p>P.S. This post is a part of DevOps Micro Internship with Agentic AI Cohort-3 by <a href="https://www.linkedin.com/in/pravin-mishra-aws-trainer/">Pravin Mishra</a>. You can start your DevOps journey by joining this <a href="https://discord.pravinmishra.com/">Discord community</a> ( <a href="https://discord.pravinmishra.com/">https://discord.pravinmishra.com/</a> ).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e56ad8a38095" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>