<?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 Naebi Sede on Medium]]></title>
        <description><![CDATA[Stories by Naebi Sede on Medium]]></description>
        <link>https://medium.com/@sede.willford?source=rss-8309172245dc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*AHSsS3OWExACa6-vKoET2A.jpeg</url>
            <title>Stories by Naebi Sede on Medium</title>
            <link>https://medium.com/@sede.willford?source=rss-8309172245dc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 10:11:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sede.willford/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[I Built an AI-Powered Phishing Detector as a Chrome Extension and I Need Your Help Breaking It]]></title>
            <link>https://medium.com/@sede.willford/i-built-an-ai-powered-phishing-detector-as-a-chrome-extension-and-i-need-your-help-breaking-it-21f20b4193ba?source=rss-8309172245dc------2</link>
            <guid isPermaLink="false">https://medium.com/p/21f20b4193ba</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[privacy]]></category>
            <dc:creator><![CDATA[Naebi Sede]]></dc:creator>
            <pubDate>Mon, 22 Jun 2026 05:35:01 GMT</pubDate>
            <atom:updated>2026-06-22T05:35:01.402Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bqRzTMljjVLv_JxjefUHkw.png" /></figure><h4>From career-changer to shipped product: how PhishGuard came together, what’s under the hood, and why I need real users to test it right now.</h4><p>Six months ago I was trading agricultural commodities. Today I just published a Chrome extension that uses a fine-tuned BERT model to catch phishing links in real time, backed by a FastAPI service, a domain whitelist of a million sites, and an optional second opinion from a fraud-detection API.</p><p>This is the story of <strong>PhishGuard, </strong>what it does, how it works, and why I’m writing this post mostly to ask you for something: <strong>please install it, try to break it, and tell me what goes wrong</strong>.</p><h4><strong>The problem I kept running into</strong></h4><p>Phishing isn’t going away. If anything, it’s getting harder to spot. Typosquatted domains, lookalike subdomains, shortened URLs, fake login pages hosted on throwaway TLDs, the tricks are well documented, but most people (myself included, on a bad day) still click without thinking twice.</p><p>Browser-native protections catch the obvious stuff. But a lot of phishing slips through simply because it’s <strong>new,</strong> not yet on anyone’s blocklist, not yet reported, not yet indexed. I wanted something that could catch threats <strong>before</strong> they’re famous.</p><p>So I built PhishGuard to scan every link on every page you visit, score it, and warn you before you click, without needing the threat to already be known.</p><h4><strong>How it actually works</strong></h4><p>PhishGuard runs detection through several layers, each one designed to catch what the previous layer might miss:</p><p><strong>1. Domain whitelist (instant, free)</strong><br>Before anything else happens, PhishGuard checks the link against the Tranco top-1-million list plus a custom whitelist. If you’re clicking something on GitHub or Wikipedia, the extension already knows it’s safe and does nothing else, no API call, no delay.</p><p><strong>2. Rule-based heuristics</strong><br>Suspicious TLDs, excessive subdomains, known phishing keywords like <strong>verify</strong> or <strong>secure-account</strong>, unencrypted HTTP, these patterns get scored before anything touches a machine learning model.</p><p><strong>3. A fine-tuned BERT model</strong><br>URLs that pass the first two filters get analyzed by a phishing-specific BERT model. This is where novel, never-seen-before phishing attempts get caught, the kind that wouldn’t be on any blocklist yet.</p><p><strong>4. Optional enhanced confirmation</strong><br>For anything flagged as suspicious, PhishGuard can optionally cross-check with a URL reputation API for a second signal, reducing false positives and confirming real threats with more confidence.</p><p>The result shows up directly in your browser: dangerous links get a red warning icon inline, the popup tells you instantly whether the page you’re on is safe, and you get a running count of what’s been caught.</p><h4><strong>Why I’m asking for help now</strong></h4><p>I tested this myself for weeks. I tested it with synthetic phishing URLs, with real flagged domains, with edge cases I tried to think up on purpose. But I am one person, on one network, with one browsing pattern. That is not nearly enough coverage for something meant to protect people from active threats.</p><h4>I need:</h4><p>- <strong>People actually using it day to day,</strong> not just on test pages, but in real inboxes, real browsing sessions, real link-clicking chaos<br>- <strong>False positives reported, </strong>if PhishGuard flags something that’s clearly safe, I want to know exactly what it was<br>- <strong>False negatives reported, </strong>if something phishy slipped through untouched, that’s the most important bug I could possibly fix<br>- <strong>Performance feedback, </strong>does it slow your browsing down? Does the popup lag? Does anything feel janky?<br>- <strong>Edge cases, </strong>weird URL structures, unusual sites, regional domains, anything that doesn’t fit a typical pattern</p><p>This is genuinely early. I built this as a solo project while studying for Security+ and applying for SOC roles, not as a funded startup with a QA team. The gap between “works on my machine” and “works for everyone” is exactly where I need outside eyes.</p><h4><strong>What I’d love from you</strong></h4><p>If you’re willing to install it and poke at it for a week:</p><p>1. Use it normally, browse like you always do.<br>2. If something gets flagged that shouldn’t be, note the URL and what happened.<br>3. If you find a phishing link that <strong>doesn’t</strong> get flagged, that’s even more valuable, tell me.<br>4. Leave a review, good or bad, with specifics if possible.<br>5. If you hit a bug, technical or visual, I want to hear about it.</p><p>I’ll be actively iterating based on what comes in. This isn’t a “ship it and walk away” project, it’s the start of something I plan to keep improving, ideally with a community of people who actually care about this problem.</p><h4><strong>Where to find it</strong></h4><p><strong>Install PhishGuard from the Chrome Web Store:</strong><br>[chromewebstore.google.com/detail/phishguard-fraud-detectio](<a href="https://chromewebstore.google.com/detail/phishguard-fraud-detectio/ledibdobihmijanhdfcedcjceccafemm">https://chromewebstore.google.com/detail/phishguard-fraud-detectio/ledibdobihmijanhdfcedcjceccafemm</a>)</p><p>It’s free to install, with an optional enhanced mode if you want a second layer of detection backed by a short free trial.</p><p>If phishing has ever cost you, almost cost you, or just annoyed you with how close it got, I’d love for you to give this a shot and tell me where it falls short.</p><h4><strong>Send me your feedback</strong></h4><p>The best place to report bugs, false positives, false negatives, or anything that feels off:<br><strong>[github.com/Naebi-stack/fraud-detection-extensions](</strong><a href="https://github.com/Naebi-stack/fraud-detection-extensions#-contact)**"><strong>https://github.com/Naebi-stack/fraud-detection-extensions#-contact)</strong></a></p><p>You can also find me here:<br>- <strong>GitHub: (</strong><a href="https://github.com/Naebi-stack"><strong>https://github.com/Naebi-stack</strong></a><strong>)</strong><br>- <strong>LinkedIn: (</strong><a href="https://www.linkedin.com/in/naebi-sede-49b38362"><strong>https://www.linkedin.com/in/naebi-sede-49b38362</strong></a><strong>)</strong></p><p>Thanks for reading, and thanks in advance to anyone willing to be an early tester. This kind of feedback loop is the only way a security tool like this actually gets better.</p><p><strong>Building in public, mistakes and all. If you find something broken, that’s not a failure — that’s the whole point of asking.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21f20b4193ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How i built a Vulnerability Assessment AI Model with Keras & Gemini Embeddings]]></title>
            <link>https://medium.com/@sede.willford/how-i-built-a-vulnerability-assessment-ai-model-with-keras-gemini-embeddings-71b9fa093777?source=rss-8309172245dc------2</link>
            <guid isPermaLink="false">https://medium.com/p/71b9fa093777</guid>
            <category><![CDATA[keras]]></category>
            <category><![CDATA[vulnerability-assessment]]></category>
            <category><![CDATA[vulnerability]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Naebi Sede]]></dc:creator>
            <pubDate>Tue, 10 Jun 2025 21:00:27 GMT</pubDate>
            <atom:updated>2025-06-10T21:00:27.174Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*my429KuNMWBN4zdynJ7CMA.jpeg" /></figure><p>Last week, I was deep in the trenches of building <strong>CySec-EmbedClassify</strong> — a deep learning initiative aimed at classifying cybersecurity-related text using language embeddings and Keras.</p><p>My initial goal? Ambitious.😎</p><p>I wanted one unified model that could classify:</p><ul><li>🔐 Risk Type</li><li>🕳️ Vulnerability Type</li><li>🧩 Assessment Focus</li></ul><p>But as I dove into the realities of preprocessing fragmented cybersecurity data, manually labeling thousands of records, and dealing with evolving taxonomies, I had to pivot.</p><blockquote><strong><em>“One model to rule them all?” Not yet.</em></strong></blockquote><h4>✂ Reframing the Architecture: One Model per Task</h4><p>Instead of creating a multi-output architecture too complex for current constraints, I split the problem into <strong>three individual models</strong> — each focused on one classification target. This not only simplified my work but also gave me:</p><ul><li>💡 Modular design and faster iterations</li><li>🧪 Easier experimentation and debugging</li><li>🧹 Greater control over label quality and accuracy</li></ul><h4><strong>✅ First Milestone: The Vulnerability Assessment Classifier</strong></h4><p>The first model I completed focuses on <strong>classifying vulnerabilities based on their textual descriptions</strong> into CWE-style categories — things like <strong>Input Validation</strong>, <strong>Improper Access Control</strong>, or <strong>Code Injection</strong>.</p><h3>🔄 The Workflow in Action</h3><h4><strong>1. Dataset Preparation</strong></h4><p>I used a cleaned vulnerability dataset from NVD sources, stored in cleaned_nvd_vulnerabilities.csv. It included thousands of labeled CVE-style descriptions tied to their corresponding CWEs.</p><h4><strong>2. Embedding with Gemini API</strong></h4><p>Each description was passed through the <strong>Google Gemini embedding API</strong>, which transformed it into a rich, high-dimensional vector — capturing semantic nuances beyond simple keyword matching.</p><h4><strong>3. Model Architecture</strong></h4><p>I built a feedfoward or straightforward Keras neural network:</p><pre>model = Sequential([<br>    Input(shape=(embedding_dim,)),<br>    Dense(256, activation=&#39;relu&#39;),<br>    Dropout(0.3),<br>    Dense(128, activation=&#39;relu&#39;),<br>    Dropout(0.2),<br>    Dense(num_classes, activation=&#39;softmax&#39;)<br>])</pre><p>It’s fast, interpretable, and surprisingly performant given the data size.</p><h4><strong>4. Training &amp; Evaluation</strong></h4><p>With the embeddings in hand and labels cleaned, I trained the model and evaluated it using precision, recall, F1 score, and confusion matrices. I was able to visualize predictions and debug misclassifications interactively.</p><h4><strong>5. Gen AI Summaries (Bonus)</strong></h4><p>To make outputs more interpretable, I added <strong>LLM-powered summaries</strong> after each prediction — giving actionable context to each predicted vulnerability type.<br> This was crucial because CWE categories can sometimes feel abstract to newcomers.</p><h3><strong>🔬 Lessons Learned</strong></h3><ul><li>🧹 <strong>Garbage in, garbage out</strong> — Label cleaning and consistency were half the battle</li><li>🛠️ <strong>Gemini embeddings</strong> were effective and fast, even with modest resources</li><li>📊 <strong>Visualizations</strong> (confusion matrix, class frequency) helped detect weak spots</li><li>💬 <strong>Generative summaries</strong> add explainability — and are helpful for analysts</li></ul><h3><strong>📌 What’s Next</strong></h3><p>With the Vulnerability Classifier done, I’m moving on to the <strong>Risk Type</strong> and <strong>Assessment Focus</strong> classifiers. Then, I plan to evaluate if model ensembling or cross-task training could improve accuracy across the board.</p><p><strong>🔗 Try It Yourself</strong></p><p>You can find the full notebook and dataset here:</p><ul><li>📁 GitHub: CySec-EmbedClassify Repo (pending)</li><li>🧪 Colab: L<a href="https://colab.research.google.com/drive/1dC85izKb9yjpi_4HxPV2BQKOhyuVByrO?usp=sharing">aunch the Vulnerability Model Notebook</a></li><li>📊 Kaggle: <a href="https://www.kaggle.com/code/naebisede/vulnerability-assessment-model-cyseembedclassify">View Dataset + Explorations</a></li></ul><p>If you’re working on applying AI to cybersecurity or threat intel, I’d love to connect. Let’s build together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71b9fa093777" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building CySec-EmbedClassify: My Journey into Gen AI-Powered Cybersecurity Text Classification]]></title>
            <link>https://medium.com/@sede.willford/building-cysec-embedclassify-my-journey-into-gen-ai-powered-cybersecurity-text-classification-6c6e169fcd1e?source=rss-8309172245dc------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c6e169fcd1e</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[vulnerability-assessment]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[risk-assessment]]></category>
            <category><![CDATA[incident-response]]></category>
            <dc:creator><![CDATA[Naebi Sede]]></dc:creator>
            <pubDate>Wed, 14 May 2025 20:42:45 GMT</pubDate>
            <atom:updated>2025-05-15T18:05:26.921Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>CySec-EmbedClassify: My Journey into Gen AI-Powered Cybersecurity Text Classification</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LCZk6Mab8YYBsX9TLNJrpQ.png" /><figcaption>CySec-EmbedClassify</figcaption></figure><p>Over a week ago, I embarked on a hands-on deep learning journey to build <strong>CySec-EmbedClassify</strong> — a Keras-powered model designed to classify cybersecurity-related text into <strong>Risk Type</strong>, <strong>Vulnerability Type</strong>, and <strong>Assessment Focus</strong>, using modern language embeddings.</p><h4><strong>The Goal</strong></h4><p>My objective was to apply <strong>generative AI techniques and text embedding models</strong> to real-world cybersecurity data — classifying unstructured text from vulnerability databases, threat intelligence reports, and risk assessments.</p><p>The idea? Build something <strong>practical</strong> that could help analysts or researchers quickly categorize and navigate critical cybersecurity information.</p><h4><strong>⏱️ Week One: Getting Started</strong></h4><p>I began by defining the scope and objectives around three classification targets:</p><ul><li><strong>Risk Type</strong></li><li><strong>Vulnerability Type</strong></li><li><strong>Assessment Focus</strong></li></ul><p>From day one, I planned to use <strong>Google Gemini embeddings</strong> for feature representation, process data using <strong>Pandas</strong>, and design the classification model with <strong>Keras</strong>.</p><p>The excitement was real — but so were the challenges.</p><h4><strong>🧱 The Biggest Hurdle: Finding Real, Usable Data</strong></h4><p>The <strong>most time-consuming challenge</strong> was <strong>data acquisition</strong>. Cybersecurity datasets, especially open-source ones, are scattered across formats and repositories. Many are unstructured, outdated, or unlabeled — which makes them tricky for training deep learning models.</p><p>I needed structured, high-volume text samples from:</p><ul><li>Vulnerability descriptions (CVEs, CWEs)</li><li>Incident reports and threat intel feeds</li><li>Risk assessment content (audit findings, controls)</li></ul><p>Finding over <strong>10,000 records for each</strong> category proved difficult.</p><h4><strong>🔧 The Solution: Aggregating Public Threat Intel from GitHub</strong></h4><p>To solve this, I wrote a custom <strong>Python downloader and extractor script</strong> that fetches .csv and .json threat intelligence data from open-source GitHub repositories like:</p><ul><li><a href="https://github.com/mitre/cti">MITRE CTI</a></li><li><a href="https://nvd.nist.gov/">NVD feeds</a></li><li><a href="https://github.com/davidonzo/Threat-Intel">DigitalSide Threat-Intel</a></li><li><a href="https://github.com/Bert-JanP/Open-Source-Threat-Intel-Feeds">Open-Source Threat Feeds</a></li></ul><p>The script recursively scanned extracted files, loaded valid datasets into <strong>Pandas DataFrames</strong>, and compiled a 10,000-record sample for experimentation.</p><p><em>Define Open-Source GitHub Repositories</em></p><pre>repos = {<br>    &quot;MITRE_CTI&quot;: &quot;https://github.com/mitre/cti/archive/refs/heads/master.zip&quot;,<br>    &quot;DigitalSide_ThreatIntel&quot;: &quot;https://github.com/davidonzo/Threat-Intel/archive/refs/heads/master.zip&quot;,<br>    &quot;OpenSourceThreatFeeds&quot;: &quot;https://github.com/Bert-JanP/Open-Source-Threat-Intel-Feeds/archive/refs/heads/master.zip&quot;<br>}</pre><p><em>Combine and Save a 10,000-Record Sample for Experimentation</em></p><pre>if dataframes:<br>    combined_df = pd.concat(dataframes, ignore_index=True)<br>    sample_df = combined_df.head(10000)<br>    sample_df.to_csv(&quot;github_threat_intel_dataset.csv&quot;, index=False)<br>    print(&quot;Saved sample to github_threat_intel_dataset.csv&quot;)<br>else:<br>    print(&quot;No usable data extracted.&quot;)</pre><p>I also explored additional resources like Kaggle datasets, CVE/CWE JSON exports, and threat feed APIs.</p><h4><strong>Where I Am Now: Preprocessing &amp; Manual Labeling</strong></h4><p>At the time of writing, I’m currently in the <strong>data cleaning and manual labeling</strong> phase — arguably the most painstaking yet critical part of the process.</p><p>This involves:</p><ul><li>Removing irrelevant fields</li><li>Filtering out noisy or low-value records</li><li>Mapping raw descriptions to consistent categories like:</li><li>Phishing</li><li>Misconfiguration</li><li>Access Control</li><li>Business Risk</li></ul><p>This step requires <strong>domain knowledge</strong> and extreme <strong>attention to detail</strong> — but it lays the groundwork for an effective model.</p><h4><strong>What’s Next: Embeddings + Keras = CySec-EmbedClassify 😜</strong></h4><p>Once preprocessing is complete, I’ll begin building the <strong>CySec-EmbedClassify</strong> model.</p><p>The plan:</p><ol><li><strong>Embed the text</strong> using <strong>Google Gemini embeddings</strong> to convert each snippet into a high-dimensional vector representation.</li><li><strong>Feed the vectors</strong> into a <strong>Keras deep learning model</strong> with:</li></ol><ul><li>Dense layers for feature learning</li><li>Dropout for regularization</li><li>Categorical output heads — one for each classification target</li></ul><p>The model will be trained on the cleaned and labeled dataset, with a focus on <strong>accuracy, robustness, and interpretability</strong>.</p><h4><strong>🎓 Lessons So Far</strong></h4><ul><li>📊 <strong>Getting clean, real-world data</strong> is the hardest and most time-consuming part of any applied AI project.</li><li>🛠️ <strong>Automation and scripting</strong> are essential when dealing with fragmented open-source cybersecurity data.</li><li>🧠 <strong>Labeling</strong> isn’t glamorous, but it’s absolutely mission-critical. It demands patience and precision. <em>God bless the creator of pandas</em>😁</li></ul><h4>📌 Coming Soon</h4><p>In the coming days, I’ll continue refining the dataset and begin experimenting with the neural architecture.</p><p>I’ll share more soon — including:</p><ul><li>First-pass model results</li><li>Code and architecture details</li><li>Lessons from model tuning and evaluation</li></ul><p>If you’re also exploring Gen AI in the cybersecurity space, stay tuned — or feel free to connect. Always happy to learn together!</p><p>🔗 Connect Explore the Project</p><p>Linkedin: <a href="http://www.linkedin.com/in/naebi-sede-49b38362">www.linkedin.com/in/naebi-sede-49b38362</a></p><p>links to Github, Kaggle, and Colab will be provided soon.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c6e169fcd1e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Built a Generative AI Cybersecurity Tool to Detect Fraudulent Websites]]></title>
            <link>https://medium.com/@sede.willford/how-i-built-a-generative-ai-cybersecurity-tool-to-detect-fraudulent-websites-115650fddb09?source=rss-8309172245dc------2</link>
            <guid isPermaLink="false">https://medium.com/p/115650fddb09</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[llm-applications]]></category>
            <category><![CDATA[osint-investigation]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cyber-security-awareness]]></category>
            <dc:creator><![CDATA[Naebi Sede]]></dc:creator>
            <pubDate>Sun, 20 Apr 2025 18:25:29 GMT</pubDate>
            <atom:updated>2025-04-21T17:05:09.908Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_xg6KvJUT4upLd9OdxxVRQ.png" /></figure><h4>A GenAI-powered capstone project using Google Gemini, VirusTotal, and community intelligence.</h4><p>💬 <em>“What if AI could analyze a suspicious URL like a human threat analyst — summarizing, comparing, and classifying it in seconds?”</em></p><p>That’s the idea behind this project. As part of the <strong>GenAI Intensive Course 2025 Capstone</strong>, I developed a tool that combines <strong>Generative AI</strong>, <strong>web scraping</strong>, and <strong>real-time threat feeds</strong> to detect and classify fraudulent websites — from phishing links to fake storefronts.</p><h3>🚨 The Problem: Detecting Online Scams is Getting Harder</h3><p>Fraudulent websites are constantly evolving — sometimes disappearing within hours, only to reappear under new domains. Traditional detection techniques like URL blacklists or manual reporting can’t scale with the speed and sophistication of modern scams.</p><h3>🧠 My Solution: A GenAI-Powered Cybersecurity Tool</h3><p>This project leverages <strong>Google’s Gemini models: </strong><em>gemini-1.5-flash-001-tuning</em>, <em>embedding-001</em>, and the <strong>VirusTotal API</strong> to analyze and flag potentially malicious websites.</p><p>Here’s what the tool does:</p><ul><li>🔗 Accepts <strong>user-reported URLs</strong></li><li>🌐 Scrapes <strong>webpage content</strong> using BeautifulSoup</li><li>🧾 Summarizes page content via <strong>Gemini </strong>using prompts</li><li>🧠 Compares to known threats using <strong>text embeddings</strong></li><li>🛡️ Scores risk level with help from <strong>VirusTotal</strong></li></ul><h3>💻 Key Implementation Snippets:</h3><h4><strong>1. Web Scraping with </strong><strong>BeautifulSoup</strong></h4><pre>import requests<br>from bs4 import BeautifulSoup<br><br>def fetch_page_text(url):<br>    response = requests.get(url)<br>    soup = BeautifulSoup(response.content, &quot;html.parser&quot;)<br>    return soup.get_text()</pre><p><em>This scrapes the raw visible text of a webpage, forming the basis for further analysis.</em></p><p><strong>2.</strong> <strong>Content Summarization with Gemini</strong></p><pre>def summarize_text(text):<br>    prompt = f&quot;Summarize this website content to detect potential fraud:\n\n{text}&quot;<br>    return gemini_model.generate_content(prompt).text</pre><p><em>Gemini summarizes the scraped content, helping highlight suspicious elements like fake login pages or urgent calls to action.</em></p><p><strong>3. Embedding &amp; Similarity Detection</strong></p><pre>from sklearn.metrics.pairwise import cosine_similarity<br><br>def get_similarity_score(embedding1, embedding2):<br>    return cosine_similarity([embedding1], [embedding2])[0][0]</pre><p><em>Compares the site’s semantic profile with known scam patterns using text embeddings.</em></p><p><strong>4. VirusTotal Integration (RAG-Style Augmentation)</strong></p><pre>import requests<br><br>def virustotal_lookup(api_key, url):<br>    headers = {&quot;x-apikey&quot;: api_key}<br>    response = requests.get(f&quot;https://www.virustotal.com/api/v3/urls/{url}&quot;, headers=headers)<br>    return response.json()</pre><p><em>While not a full RAG pipeline, this step augments the Gemini analysis by retrieving real-time threat data — a </em><strong><em>lightweight form of Retrieval-Augmented Generation (RAG)</em></strong><em>.</em></p><p><strong>5. 📄 Final Report Generation</strong></p><pre>def generate_report(url, summary, classification, similarity_score, vt_result):<br>    return {<br>        &quot;URL&quot;: url,<br>        &quot;Summary&quot;: summary,<br>        &quot;Classification&quot;: classification,<br>        &quot;Similarity Score&quot;: round(similarity_score, 2),<br>        &quot;VirusTotal Result&quot;: vt_result,<br>    }</pre><p><em>This produces a compact, human-readable report combining AI insights and external threat intelligence.</em></p><p><strong>6. Final Scoring and Explanation</strong></p><p>Results include:</p><ul><li>✅ Threat classification (e.g. phishing, scam)</li><li>📄 Gemini-generated explanation</li><li>🧬 Similarity score</li><li>🧠 VirusTotal threat intelligence</li></ul><h3>What GenAI Skills Did I Apply?</h3><p>This project showcases<strong> four</strong> <strong>key Generative AI techniques</strong>:</p><ol><li><strong>Content Summarization</strong> using LLMs</li><li><strong>Text Embedding Generation</strong> for similarity matching</li><li><strong>Prompt Engineering</strong> for threat analysis and classification</li><li><strong>RAG-Style Augmentation </strong>for <em>retrieving real-time threat data</em></li></ol><h3>Limitations</h3><ul><li>Pages that rely heavily on JavaScript can’t be scraped (yet).</li><li>Gemini models have token limits — large sites are truncated or chunked.</li><li>The fraud dataset is relatively small, affecting precision.</li></ul><h3>🧭 What’s Next?</h3><p>Here’s where I plan to take it:</p><ul><li>Build a <strong>Chrome extension</strong> to flag malicious links in real time</li><li>Use <strong>crowd feedback</strong> to improve classification</li><li>Add <strong>dynamic content scraping</strong> via Selenium</li><li>Create a <strong>threat dashboard</strong> with automated summaries, making it useful for OSINT</li><li>Making the <strong>community response</strong> more functional</li></ul><h3>Final Thoughts</h3><p>This capstone project let me blend my cybersecurity passion with real-world AI tools to build something useful. By fusing community intelligence with powerful generative models, we can make the internet just a bit safer.</p><blockquote><em>🙌 Thanks to </em><strong><em>Google</em></strong><em>, </em><strong><em>Kaggle</em></strong><em>, and the </em><strong><em>GenAI Intensive Course 2025</em></strong><em> for this opportunity.</em></blockquote><h3>Want to See the Code?</h3><p>Check out the full Kaggle notebook here:<br> 🔗 <em>[</em><a href="https://www.kaggle.com/code/naebisede/capstone-project-5-day-genai-2025">https://www.kaggle.com/code/naebisede/capstone-project-5-day-genai-2025</a><em>]</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=115650fddb09" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>