<?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 Cheng Yee L. on Medium]]></title>
        <description><![CDATA[Stories by Cheng Yee L. on Medium]]></description>
        <link>https://medium.com/@shadowandy?source=rss-f1d19be98c59------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pzwgQ3WXklEYcIGR4Sbn-A.jpeg</url>
            <title>Stories by Cheng Yee L. on Medium</title>
            <link>https://medium.com/@shadowandy?source=rss-f1d19be98c59------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 02:32:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shadowandy/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[From Days to Hours: Building a Deterministic IM8 Controls Agent Skill with Agentic Engineering]]></title>
            <link>https://shadowandy.medium.com/from-days-to-hours-building-a-deterministic-im8-controls-agent-skill-with-agentic-engineering-d1fa95228a8f?source=rss-f1d19be98c59------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1fa95228a8f</guid>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Cheng Yee L.]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 14:40:57 GMT</pubDate>
            <atom:updated>2026-09-18T23:40:23.965Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bJVWGOLpkQj3O34hSiBycw.jpeg" /><figcaption>Deterministic agent skill to query IM8 controls. (image created using Gemini)</figcaption></figure><p>Every architect and engineer working on Singapore public sector digital projects eventually faces the same compliance bottleneck: navigating the <a href="https://info.standards.tech.gov.sg/?utm_source=gemini">Singapore Government ICT&amp;SS Policy Reform (IM8)</a>.</p><p>Architecting a compliant digital service requires cross-referencing baseline System Security Plans (SSPs), checking control applicability across Low-, Medium-, or High-Risk Cloud profiles, and identifying supplementary controls whenever emerging technologies like Generative AI enter the solution architecture. Manually searching and reading through documentation portals for every architectural variation imposes a heavy cognitive tax.</p><p>Turning to general-purpose conversational LLMs like ChatGPT, Claude, or Gemini is an intuitive alternative. However, applying out-of-the-box conversational agents to compliance workflows quickly reveals significant reliability and accuracy limitations.</p><p>To solve this, I built a deterministic agent skill. You can find the complete source code for this project in my GitHub repository: <a href="https://github.com/shadowandy/sg-im8-reform-skill"><strong>shadowandy/sg-im8-reform-skill</strong></a>.</p><h3>Why Web Search and Ad-Hoc Scraping Fall Short</h3><p>The initial, straightforward approach was to equip an agent with its default web search or browsing tool directed at info.standards.tech.gov.sg.</p><p>In practice, dynamic web retrieval introduces friction that regulatory compliance workflows cannot tolerate:</p><ul><li><strong>Non-Deterministic Retrieval:</strong> Web search tools return variable page snippets depending on how the model phrases its underlying search queries. A subtle shift in prompt wording can cause an agent to retrieve high-level introductory pages rather than specific control clauses.</li><li><strong>Context Bloat vs. Incomplete Scope:</strong> Fetching raw HTML either bloats the context window with navigation shells and scripts or truncates essential baseline lists.</li><li><strong>Hallucinated Baselines:</strong> When web retrieval provides incomplete context, language models tend to bridge the missing links by inventing plausible-sounding controls that have no grounding in actual policy text.</li><li><strong>Failure to Aggregate Complex Profiles:</strong> Consider a realistic query: <em>“What baseline controls apply to a Medium-Risk Cloud system incorporating an LLM?”</em> Answering this accurately requires retrieving and intersecting two separate policy tracks — the base cloud profile and the GenAI supplementary controls. A basic web-search agent rarely performs multi-hop profile discovery and synthesis reliably.</li></ul><p>Compliance engineering demands exact, deterministic references. Factual retrieval must be grounded in an authoritative source, reserving the LLM’s strengths for natural-language intent interpretation, cross-profile reasoning, and conversational synthesis.</p><h3>The Architecture: Decoupled Engineering Tracks</h3><p>To eliminate probabilistic retrieval errors, I packaged the solution as a dedicated <strong>Agent Skill</strong> backed by a localized, structured dataset and strict execution boundaries.</p><p>Instead of relying on a monolithic agent, the system separates concerns across three distinct engineering tracks:</p><ul><li><strong>Track 1: Procedural Ingestion and Normalization (Python):</strong> A deterministic Python pipeline (extract_oscal.py) ingests authoritative policy documents from the GovTech portal and normalizes them into token-optimized JSON catalogs aligned with Open Security Controls Assessment Language (OSCAL) concepts. This maps control IDs, objective statements, and baseline levels (0, 1, 2) into clean key-value structures.</li><li><strong>Track 2: Packaging Automation (Bash):</strong> A lightweight shell automation script (package-skill.sh) packages the normalized catalogs, system security profiles, and skill manifests into a distributable skill bundle ready for deployment.</li><li><strong>Track 3: The Grounded Agent Skill (Runtime LLM):</strong> Rather than running ad-hoc web searches, the agent skill executes deterministic file reads (im8.py) against the local JSON catalogs. The prompt instructions (SKILL.md) guide the LLM to route user intents to the appropriate local profile, fetch exact control clauses, and synthesize the output.</li></ul><h3>Core Engineering Principles</h3><p>Structuring this capability highlighted three critical software engineering patterns for domain-specific agent development:</p><ul><li><strong>Separation of Factual Truth and Architectural Judgment:</strong> Control wording and baseline levels are immutable facts that must be cited verbatim. System applicability — such as deciding whether an API gateway counts as a sandbox boundary — involves organizational context. The skill is explicitly instructed to present applicability mappings as draft assessments for human confirmation rather than authoritative mandates.</li><li><strong>Instruction-Led Cognitive Routing:</strong> Rather than introducing a complex external routing microservice, routing logic is governed through prompt-based heuristics in SKILL.md. The skill instructs the model to evaluate the user&#39;s architectural context (e.g., hosting type, data classification, GenAI components) and determine exactly which local catalog nodes to inspect.</li><li><strong>Selective Context Loading:</strong> Indiscriminately loading the entire control catalog into the conversation context degrades response latency and pollutes attention. Because the data is split into isolated files (gen-ai.json, medium-risk-cloud.json, etc.), the skill directs the LLM to inspect only the profiles relevant to the matched system classification.</li></ul><h3>Securing the Pipeline: CI/CD and Behavioral Verification</h3><p>Moving fast with agentic development tools like Claude Code requires establishing clear quality gates. Because this skill handles compliance data, testing was split into two distinct tracks: <strong>deterministic automation</strong> for the data pipeline, and <strong>behavioral evaluation</strong> for the agent.</p><p><strong>1. Automated CI/CD for the Data Pipeline</strong> We cannot afford parsing errors or malformed references in the underlying IM8 control data. The repository’s continuous integration pipeline (checks.yml) strictly enforces the integrity of the ingestion and packaging scripts:</p><ul><li><strong>Script Verification:</strong> Automated runners execute check.sh on every push to ensure the Python parsing logic and shell packaging scripts execute cleanly without unhandled exceptions.</li><li><strong>Data Integrity Validation:</strong> A dedicated validation script (check_data.py) acts as a quality gate, asserting that the generated JSON catalogs contain the required structural keys (control_id, statement, level) and that all baseline mappings resolve correctly before a skill bundle is packaged.</li></ul><p><strong>2. Structured Behavioral Evaluation for the Agent</strong> Because LLM reasoning is probabilistic, evaluating the agent’s cognitive routing and synthesis was conducted against a strict, manual benchmark rubric rather than standard unit tests. The skill was iteratively validated against edge cases to ensure boundary enforcement:</p><ul><li><strong>Exact Factual Lookup<br></strong><em>Query Example:</em> “What is GA-3?”<br><em>Validation Objective:</em> Verify exact control wording and level without hallucinated additions.</li><li><strong>Profile Selection<br></strong><em>Query Example:</em> “What SSP applies to a sandbox LLM?”<br><em>Validation Objective:</em> Ensure the agent correctly identifies specialized profiles.</li><li><strong>Multi-Profile Synthesis<br></strong><em>Query Example:</em> “What applies to Medium-Risk Cloud using an LLM?” <br><em>Validation Objective:</em> Test the agent’s ability to aggregate base SSP controls with GenAI controls.</li><li><strong>Ambiguity Handling<br></strong><em>Query Example:</em> [Incomplete architecture description]<br><em>Validation Objective:</em> Verify the agent surfaces missing assumptions and prompts for clarification.</li><li><strong>Hallucination Resistance</strong> <br><em>Query Example:</em> [Unsupported policy query] <br><em>Validation Objective:</em> Confirm the agent explicitly states when a topic is outside the IM8 scope.</li></ul><h3>Business Impact: Days to Hours</h3><p>Adopting an agentic development workflow fundamentally altered the delivery lifecycle of this capability:</p><ul><li><strong>Rapid Implementation Cycles:</strong> Writing bespoke scrapers, defining data schemas, structuring prompt guardrails, and setting up packaging scripts typically demands several days of manual scaffolding. Paired with Claude Code, the initial end-to-end implementation was completed, verified, and operational in <strong>hours</strong>.</li><li><strong>Enterprise-Ready Deployment:</strong> Packaged cleanly as an agent skill, the capability can be deployed into an <strong>Enterprise ChatGPT environment</strong>,<strong> Claude Code</strong>,<strong> Antigravity</strong>,<strong> etc. </strong>moving from an engineer’s experimental script to a shared, reusable enterprise capability accessible to delivery teams.</li><li><strong>Instant Regulatory Clarity:</strong> Product teams and systems architects can now interrogate multi-layered ICT&amp;SS requirements interactively, receiving reliable, grounded compliance baselines within seconds instead of combing through web portals.</li></ul><p>By replacing brittle web browsing with deterministic local data structures and bounded cognitive reasoning, compliance tooling evolves from unreliable chat experiments into dependable, production-grade engineering assets.</p><p><em>Check out the full implementation and repository at </em><a href="https://github.com/shadowandy/sg-im8-reform-skill"><em>https://github.com/shadowandy/sg-im8-reform-skill</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1fa95228a8f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Sync Your Kobo e-Reader Wirelessly with Calibre-Web and Nginx Proxy Manager]]></title>
            <link>https://shadowandy.medium.com/how-to-sync-your-kobo-e-reader-wirelessly-with-calibre-web-and-nginx-proxy-manager-ca594d4254c4?source=rss-f1d19be98c59------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca594d4254c4</guid>
            <category><![CDATA[self-hosted]]></category>
            <category><![CDATA[calibre]]></category>
            <category><![CDATA[homelab]]></category>
            <category><![CDATA[nginx]]></category>
            <dc:creator><![CDATA[Cheng Yee L.]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 06:35:13 GMT</pubDate>
            <atom:updated>2026-09-05T06:35:13.001Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0fX_9kgUKn3v5QlykwDEzQ.jpeg" /><figcaption>Syncing Kobo e-Reader with Calibre (image made with Gemini)</figcaption></figure><h3>Ditch the USB cable and sync your self-hosted ebook library over the air — securely behind an HTTPS reverse proxy.</h3><h3>Why Wire When You Can Sync Over the Air?</h3><p>While Kobo devices are widely praised across the reading and homelab communities for being far more flexible than closed ecosystems like Amazon Kindle, their native cloud sync remains tethered to Kobo’s retail catalog. Readers who self-host extensive digital collections using Calibre are typically relegated to manual USB sideloading: digging out a cable, plugging into a PC, and dragging files over one by one.</p><p>By pairing your device with <strong>Calibre-Web’s Kobo Sync API</strong>, you can retarget your e-reader to sync directly with your self-hosted library. You get automatic collection/shelf syncing, reading progress tracking, and effortless over-the-air book delivery — all while retaining full access to OverDrive, Pocket, Dropbox, and the native Kobo Store.</p><h3>The Technical Hurdle</h3><p>Calibre-Web’s internal Kobo Sync mechanism constructs its book download payload using plain http:// endpoints. When exposing your setup securely to the public internet using an HTTPS reverse proxy (port 443), this protocol mismatch breaks downloads on the e-reader.</p><p>In this guide, we will configure <strong>Nginx Proxy Manager (NPM)</strong> with response-body rewriting to handle SSL termination cleanly, configure <strong>Calibre-Web</strong>, and point your <strong>Kobo</strong> at your private cloud endpoint.</p><h3>Prerequisites</h3><p>Before starting, make sure you have:</p><ul><li><strong>A Kobo e-reader</strong> (e.g., Kobo Clara, Libra Colour, Sage, Elipsa).</li><li><strong>A computer</strong> and a <strong>USB data cable</strong> (needed only once to edit a configuration file).</li><li><strong>Calibre-Web</strong> running with your EPUB library.</li><li><strong>Nginx Proxy Manager (NPM)</strong> configured with a public domain or subdomain and a valid SSL certificate (e.g., Let’s Encrypt).</li></ul><blockquote><strong><em>Note on File Formats:</em></strong><em> Kobo’s native sync mechanism only supports </em><strong><em>EPUB</em></strong><em> and </em><strong><em>KEPUB</em></strong><em> formats. Other formats (like PDF or CBR/CBZ) will not sync over the air and must still be sideloaded manually.</em></blockquote><h3>Step 1: Configure Nginx Proxy Manager (NPM)</h3><p>Because Calibre-Web returns book download URLs formatted as http://, we use Nginx’s sub_filter module to intercept the JSON API response and dynamically rewrite those links to https:// before sending them to your Kobo.</p><ol><li>Open <strong>Nginx Proxy Manager</strong> and edit the <strong>Proxy Host</strong> for your Calibre-Web instance.</li><li>Ensure your SSL certificate is attached and <strong>Force SSL</strong> is enabled.</li><li>Navigate to the <strong>Advanced</strong> tab and paste the following custom configuration into the <strong>Custom Nginx Configuration</strong> block:</li></ol><p>Nginx</p><pre>proxy_set_header Host $http_host;<br>proxy_set_header Upgrade $http_upgrade;<br>proxy_set_header Connection &quot;upgrade&quot;;<br>proxy_set_header X-Forwarded-Host $http_host;<br>proxy_set_header X-Forwarded-Proto $scheme;<br>proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;<br>proxy_set_header X-Scheme $scheme;<br>proxy_set_header X-Script-Name //;<br>proxy_cookie_path / &quot;/; Secure&quot;;<br>proxy_busy_buffers_size 1024k;<br>proxy_buffers 4 512k;<br>proxy_buffer_size 1024k;<br>client_max_body_size 50M;<br><br># Response body rewrite: convert `DownloadUrls` from `http://` to `https://`<br>sub_filter_types application/json;<br>sub_filter &#39;http://&#39; &#39;https://&#39;;<br>sub_filter_once off;</pre><h3>Step 2: Configure Calibre-Web</h3><p>Next, we enable Kobo sync globally and generate a personal sync token for your user account.</p><h3>2.1 Enable the Kobo Sync Feature</h3><ul><li>Log in to Calibre-Web as an administrator.</li><li>Navigate to <strong>Admin</strong> &gt; <strong>Edit Basic Configuration</strong> &gt; <strong>Feature Configuration</strong>.</li><li>Check <strong>Enable Kobo Sync</strong>.</li><li>Check <strong>Proxy unknown requests to Kobo Store</strong> (ensures your official store browsing and syncing continue to work normally).</li><li>For <strong>Server External Port (for port forwarded API calls)</strong>, specify 80.</li><li><em>(Recommended)</em> Check <strong>Embed Metadata to Ebook File on Download/Conversion/e-mail</strong>.</li><li>Click <strong>Save</strong>.</li></ul><h3>2.2 Verify Kepubify Conversion</h3><p>To take full advantage of Kobo-specific features (like fast page turns, reading statistics, and chapter time estimates), Calibre-Web can convert EPUBs to KEPUB on the fly using kepubify.</p><ul><li>In <strong>Admin</strong> &gt; <strong>Edit Basic Configuration</strong>, go to <strong>External Binaries</strong>.</li><li>Ensure the <strong>Path to Kepubify E-Book Converter</strong> points to your binary (typically /usr/bin/kepubify or /usr/bin).</li><li>Click <strong>Save</strong>.</li></ul><h3>2.3 Generate Your Kobo Sync Token</h3><ul><li>Login to Calibre-Web.</li><li>Click on your <strong>User Profile</strong> (your username in the top navigation bar).</li><li>Check <strong>Sync only books in selected shelves with Kobo</strong> (this gives you full control over which books land on your device).</li><li>Under <strong>Kobo Sync Token</strong>, click <strong>Create/View</strong>.</li><li>Copy the generated URL string. It will look like this:</li></ul><pre>api_endpoint=http://yourdomain.com/kobo/abcdef123456</pre><p><em>(Keep this handy — we will change </em>http://<em> to </em>https://<em> in the next step).</em></p><h3>2.4 Create a Sync Shelf &amp; Add Books</h3><ul><li>Login to Calibre-Web.</li><li>On the left sidebar, click <strong>Create a Shelf</strong>.</li><li>Name your shelf (e.g., Kobo Shelf).</li><li>Check <strong>Sync this shelf with Kobo device</strong>, then click <strong>Save</strong>.</li><li>Browse your library, open any book title’s details, click <strong>Add to shelf</strong>, and select your newly created shelf.</li></ul><h3>Step 3: Configure the Kobo e-Reader</h3><p>Now we direct the Kobo hardware to talk to your Calibre-Web API endpoint instead of the standard Kobo servers.</p><h3>3.1 Edit Kobo eReader.conf</h3><ul><li>Connect your Kobo to your computer using the USB data cable and tap <strong>Connect</strong> on the e-reader screen.</li><li>Open your file manager and navigate to the root directory of the Kobo.</li><li>Open the .kobo/Kobo folder. <em>(Note: On macOS and Linux, folders starting with a dot </em>.<em> are hidden by default; press </em>Cmd + Shift + .<em> on Mac or </em>Ctrl + H<em> on Linux to reveal them).</em></li><li><strong>Make a backup:</strong> Copy Kobo eReader.conf and rename the copy to Kobo eReader.conf.backup.</li><li>Open Kobo eReader.conf in a plain text editor (e.g., VS Code, Notepad++, or TextEdit in plain text mode).</li><li>Locate the section header:</li></ul><pre>[OneStoreServices]</pre><ul><li>Modify or replace the following lines. <strong>Crucial:</strong> Ensure the protocol is set to <a href="https://:">https://:</a></li><li>Replace the line starting with api_endpoint:</li></ul><pre>api_endpoint=https://yourdomain.com/kobo/your-token-here</pre><ul><li>Replace the image_host line:</li></ul><pre>image_host=https://yourdomain.com/kobo/your-token-here</pre><ul><li>Locate image_url_quality_template and replace https://cdn.kobo.com/book-images with your sync URL:</li></ul><pre>image_url_quality_template=https://yourdomain.com/kobo/your-token-here/{ImageId}/{Width}/{Height}/{Quality}/{IsGreyscale}/image.jpg</pre><ul><li>Locate image_url_template and replace https://cdn.kobo.com/book-images with your sync URL:</li></ul><pre>image_url_template=https://yourdomain.com/kobo/your-token-here/{ImageId}/{Width}/{Height}/false/image.jpg</pre><ul><li>Save the file and safely <strong>Eject / Unmount</strong> the e-reader before disconnecting the USB cable.</li></ul><h3>Step 4: Test the Sync</h3><ul><li>On your Kobo, ensure Wi-Fi is connected.</li><li>Tap the <strong>Sync icon</strong> in the top navigation bar.</li><li>Your Kobo will perform its initial handshake. Within moments, you will see a collection created matching your shelf name (e.g., <em>Kobo Shelf</em>), and the EPUB/KEPUB files will download automatically.</li><li>Open the Kobo Store or OverDrive — you’ll find that native device features continue to work without a hitch.</li></ul><h3>Troubleshooting &amp; Reverting</h3><ul><li><strong>Books appear in library but won’t download:</strong> Double-check your Nginx Proxy Manager configuration. If the sub_filter directives are missing, the Kobo may try to reach internal http:// URLs that your browser or firewall blocks.</li><li><strong>Cover images are missing:</strong> Ensure the image_url_quality_template and image_url_template parameters in Kobo eReader.conf have been updated with https:// and point to the base of your token endpoint.</li><li><strong>How to revert to factory defaults:</strong> If you ever want to return the Kobo to its stock syncing configuration, connect via USB, delete Kobo eReader.conf, and rename Kobo eReader.conf.backup back to Kobo eReader.conf.</li></ul><p>Check it out here if you prefer it in <a href="https://gist.github.com/shadowandy/0112323b47d76ec77a58e769ac1efaf9">GitHub Gist</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca594d4254c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Passed the OSCP on My First Attempt (While Working a Non-Cyber Job)]]></title>
            <link>https://shadowandy.medium.com/how-i-passed-the-oscp-on-my-first-attempt-while-working-a-non-cyber-job-00d832d9e399?source=rss-f1d19be98c59------2</link>
            <guid isPermaLink="false">https://medium.com/p/00d832d9e399</guid>
            <category><![CDATA[oscp]]></category>
            <category><![CDATA[offensive-security]]></category>
            <category><![CDATA[penetration-testing]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Cheng Yee L.]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 07:11:01 GMT</pubDate>
            <atom:updated>2026-09-05T06:41:05.060Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DsV6Uiuw5jiD3_umCvopoA.png" /><figcaption>Studying almost every evening for the OSCP (image created using Gemini)</figcaption></figure><p>In June 2023, I sat for the Offensive Security Certified Professional (OSCP) exam. I kicked off at 10:30 AM, compromised all six machines by 11:00 PM, and submitted my final report by 2:30 PM the following day.</p><p>Two days later, the confirmation email landed: <strong>I had passed on my first attempt with full points on all six targets plus 10 bonus points in the bag.</strong></p><p>For context, cybersecurity is not my primary day job. If you are juggling a full-time role, family commitments, and a ticking subscription clock, here is the honest breakdown of how I prepared, handled real-time exam crises, and crossed the finish line.</p><p><strong>The 600-Hour Preparation Journey</strong></p><p>I enrolled in PEN-200 via the Learn One subscription in October 2022. Early on, work tenders and procrastination slowed my momentum down to around 4 hours a week.</p><p>The panic hit in February 2023: I had less than five months left before my subscription expired. I overhauled my schedule and committed to a strict study regimen:</p><ul><li><strong>Oct 2022 — Jan 2023:</strong> ~4 hours/ week (casual pace)</li><li><strong>Feb — May 2023:</strong> ~22 hours/ week (6 hours on weekdays, 16 hours on weekends)</li><li><strong>1 — 11 June 2023:</strong> ~40 hours/ week (ramping up)</li><li><strong>12 — 18 June 2023:</strong> ~70 hours</li><li><strong>19–20 June 2023:</strong> ~24 hours (final review)</li></ul><p>In total, I poured over <strong>600 hours</strong> into preparation, completing the 2023 challenge labs alongside module exercises to secure the critical 10 bonus points.</p><p><strong>Why Reading Walkthroughs Accelerated My Learning</strong></p><p>Hands-on practice is essential, but doing every box from scratch can be a bottleneck. I committed to hands-on boxes daily on Proving Grounds Practice, Hack The Box, and VulnHub, but I supplemented this heavily with written walkthroughs.</p><p>Reading walkthroughs allowed me to dissect an entire box in under 10 minutes and absorb different mindsets. Key takeaways I picked up purely from other people’s write-ups included:</p><ul><li>Catching hidden cron jobs efficiently using pspy.</li><li>Leveraging chisel client for reverse port mapping.</li><li>Exploiting Active Directory environments via SharpGPOAbuse.</li><li>Understanding what <em>not</em> to use (like retiring EternalBlue from modern exam tooling).</li></ul><p><strong>Exam Day Disasters: Infrastructure &amp; Connectivity</strong></p><p>My exam was scheduled for 10:00 AM (SGT). That was when Murphy’s Law kicked in.</p><ul><li><strong>WebRTC &amp; Proctoring Failures:</strong> I had skipped booking an exam infrastructure test session. My home connection repeatedly dropped the WebRTC screen share every few minutes. I had to switch immediately to my 4G mobile hotspot, introducing higher latency and lower bandwidth.</li><li><strong>OpenVPN Bridge Conflict:</strong> When the exam pack arrived at 10:23 AM, my Kali VM refused to connect to the exam VPN. Because I had swapped network adapters to mobile data, my VM’s bridged networking broke. Reconfiguring the virtual adapter to bridge directly to the hotspot fixed it just in time to start at 10:30 AM.</li></ul><p><em>Tip: Always schedule the test proctoring session beforehand, and verify VM adapter settings if you switch network sources.</em></p><p><strong>The Exam Breakdown</strong></p><p><strong>Active Directory Set (Hours 1–3)</strong></p><p>Adrenaline pushed me to enumerate aggressively. I started nmap scans on the standalone targets while focusing the first half of the day on the AD chain. The AD client responded sluggishly over the mobile connection, so I reverted the machine immediately. By hour three, I had full Domain Admin credentials.</p><p><strong>Standalone Targets &amp; The Power of Reverts (Hours 4–12)</strong></p><p>With AD down, I sustained energy on bananas and light snacks to avoid afternoon brain fog. Over mobile data, enumeration scans (nmap, gobuster, nikto, enum4linux) took longer, so I used the waiting windows to double-check my AD exploitation notes.</p><p>When a foothold exploit consistently failed on a standalone machine despite looking rock-solid, I paused and triggered a machine revert. The exploit landed cleanly on the fresh instance. When in doubt, revert.</p><p><strong>Wrapping Up &amp; Submission</strong></p><p>By late evening, I had proofs across all 6 targets (11 proofs total: standalone + AD chain).</p><ul><li><strong>Proof Verification:</strong> Before disconnecting, I systematically reconciled every submitted hash in the exam portal against my Joplin notes: IP for IP, flag for flag.</li><li><strong>Ending the Proctor Session:</strong> I formally concluded the proctored exam around 11:00 PM to get out from under the webcam monitoring and ate my first proper meal of the day.</li><li><strong>Reporting:</strong> I drafted the report until 4:00 AM, slept, performed a final review the next morning, and submitted the PDF at 2:30 PM.</li></ul><p>Passing the OSCP required deliberate time management, methodical note-taking, and keeping a cool head when technical setups go sideways.</p><p><strong>Why an Architect Went After the OSCP</strong><br>In my day-to-day role as a systems architect, my focus centers on guiding projects and product teams to build platforms that are resilient, scalable, maintainable, and secure. Much of my time is spent bridging high-level organizational strategy with technical reality: evaluating trade-offs across complex technology landscapes, establishing robust design guardrails, and aligning distributed components into a coherent, future-proof ecosystem.<br>Because dedicated penetration testing is not my primary craft, pursuing the OSCP was never about switching careers. Instead, it was about stress-testing my own foundational knowledge and understanding the attacker’s mindset at a granular level. Earning this certification gave me the grounded confidence to spot architectural security flaws early, ask sharper questions during design reviews, and guide engineering teams toward building genuinely hardened, battle-ready systems. Moving forward, I will continue leveraging community-driven challenge labs to keep those technical instincts sharp.</p><p>Check out my <a href="https://credentials.offsec.com/824c5233-eccf-4bd5-83d8-849eeba8e23f#acc.fjLVziPp">OSCP accreditation</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=00d832d9e399" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>