<?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 Anant Wairagade on Medium]]></title>
        <description><![CDATA[Stories by Anant Wairagade on Medium]]></description>
        <link>https://medium.com/@anantrw?source=rss-7999a8b5ed7d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xHH7SHR9oiCczNneQpdP3Q.jpeg</url>
            <title>Stories by Anant Wairagade on Medium</title>
            <link>https://medium.com/@anantrw?source=rss-7999a8b5ed7d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:42:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anantrw/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[Best Practices for API Security]]></title>
            <link>https://medium.com/design-bootcamp/best-practices-for-api-security-7852029abfb6?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/7852029abfb6</guid>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Tue, 25 Nov 2025 04:06:37 GMT</pubDate>
            <atom:updated>2025-11-25T12:22:41.189Z</atom:updated>
            <content:encoded><![CDATA[<p>APIs power everything from the simple apps on your phone to complex systems in the cloud. They are the vital links that enable applications, services, and devices to talk to each other and share information effortlessly. Chances are, if you’ve ever used the internet to book a plane ticket, transfer money, or even order pizza from your favorite restaurant, you’ve directly benefited from the silent work of APIs.</p><p>But here’s the thing: every API is like an open door, and might act as a potential entry point for attackers. Without a solid security framework, you are leaving that door wide open.</p><p>Fortunately, Identity and Access Management (IAM) can help you solve this. In this article, we’ll look at how it helps you ensure that only the right apps, users, and systems can access what they are supposed to — nothing more, nothing less.</p><h3>Why You Should Care About API Security</h3><p>APIs are behind most modern software integrations. They enable developers to hook up new or old software with other applications without developing the interfaces themselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*i5MZqJXMxLvEzm66Ea3i_w.png" /></figure><p>However, this increased connectivity comes at a cost as it often exposes security vulnerabilities that malicious actors can exploit. And considering that <a href="https://boomi.com/blog/api-strategy-vs-hype/#:~:text=With%20a%20new%20year%20upon,important%20and%20strategic%20in%202025.">83% of internet traffic is API traffic,</a> having a proper IAM isn’t a “nice-to-have” incentive but an utmost priority. Why? A good IAM strategy helps you:</p><p>● Know the identity of who (or what application) is calling your API.</p><p>● Have the final say on what they can do by keeping things on a need-to-know basis.</p><p>● Monitor and maintain detailed logs of all API interactions, which are crucial for identifying anomalies and meeting regulatory requirements.</p><h3>What Happens Without It?</h3><p><strong>Credential Stuffing:</strong> attackers try every stolen username and password combo they have, and without IAM, they might get lucky and walk right in.</p><p><strong>Broken Object Level Authorization (BOLA):</strong> This is a fancy way of saying someone can sneak a peek at what they shouldn’t, maybe by just changing a single digit in a web address.</p><p><strong>Man-in-the-middle (MITM) attacks</strong> mean exactly as they sound: someone is trying to listen in on or even tamper with the conversations between APIs and their users.</p><p><strong>Data breaches:</strong> This is the absolute worst-case scenario: people with ill intentions get access to what they shouldn’t have — your data.</p><h3>IAM as Gatekeeper</h3><p>IAM in the context of API, can be broken down into two core components.</p><h3>a. Authentication: Who are you?</h3><p>Before an API let’s anyone in, their identity needs to be confirmed. Here is how it is done:</p><h4>API Keys</h4><p>It’s one of the simplest forms of authentication, and they are unique strings that identify a particular user or application. These keys are issued by the API provider to the client, enabling them to monitor usage and control access. It usually accompanies every request as part of the query string in the request header or within a cookie.</p><p>While simple to implement, they usually offer limited control over who can use them, hence they are secure. That’s why they are normally used for non-critical APIs where the risk of exposure is minimal</p><h4>OAuth 2.0 &amp; OpenID Connect (OIDC)</h4><p>This is the most widely used authentication for APIs today. It allows users to authenticate via third-party services like Google and Facebook without having to share their credentials with the API itself. It introduces the concept of an access token, which can be shared with limited privileges and an expiry period.</p><p>OpenID Connect makes it more secure by adding an identity layer that provides a standardized way for clients to confirm the end-user’s identity.</p><h4>JWT (JSON Web Tokens)</h4><p>These are compact, URL-safe, and self-contained tokens often used to transmit information between parties as a JSON object. All the information the API needs is usually included within the token, making it perfect for APIs that don’t need to track every user’s session.</p><p>You can think of them as super-efficient ID cards that don’t reveal who you are but clearly state what you do based on claims. And they work perfectly as long as they are correctly validated.</p><h4>Mutual TLS (mTLS)</h4><p>This is the double handshake for top security, meaning it requires both the client and server to provide and substantiate cryptographic certificates to each other. If all checks out, they exchange data over the encrypted connection.</p><p>While this extra handshake adds some complexities, it ensures that both parties are trusted. Making it ideal for highly regulated industries like finance and healthcare APIs</p><h3>b. Authorization: What can you do?</h3><p>Once the identity checks out, you need to know what actions or resources the user can access or manipulate within the API. This requires strong authorization mechanisms such as:</p><h4>Role-Based Access Control (RBAC)</h4><p>This is the simplest way to handle access where users are assigned specific roles with predefined permissions, such as admin, editor, and viewer. It’s pretty straightforward to configure at first, but it can get messy with complex APIs that have tons of roles or confusing overlaps.</p><h4>Attribute-Based Access Control (ABAC)</h4><p>This is a more flexible approach that makes decisions based on dynamic info associated with the user (e.g., department, security clearance), the resource (e.g., data sensitivity, resource owner), the action to be performed (e.g., read, write, delete), and environmental factors (e.g., time, device, location) instead of role. This makes it more granular than RBAC.</p><h4>Policy-Based Access Control (PBAC)</h4><p>This superior authorization strategy simplifies access control complexities by combining the benefits of RBAC and ABAC. Its policies, which are written in a human-readable format, can be centralized and evaluated based on API requests. And since they are easily understood, they can be precise, such as “deny access outside office hours”.</p><p>They are popular in huge companies with a complex setup that requires precision control.</p><h3>Best Practices to Keep Your APIs Safe</h3><p>Here are the top 6 practices to keep your APIs safe:</p><h3>1. Use OAuth 2.0 + OIDC for public-facing APIs</h3><p>For any API that the public can use, always lean on OAuth 2.0 and OpenID Connect for authentication and authorization.</p><p>Avoid using simple API keys for endpoints handling sensitive data, as they are less flexible and secure.</p><h3>2. Rate Limiting and Throttling</h3><p>Implement robust rate limiting and throttling on all API endpoints to control the number of requests users make in each period.</p><p>This mitigates the risk of brute-force attacks, denial-of-service (DoS), and distributed denial-of-service (DDoS) attacks.</p><h3>3. Follow the Least Privilege Principle</h3><p>Grant the minimum necessary permissions for any user, application, or service accessing your APIs to share as little data as possible.</p><h3>4. Use Short-Lived tokens + Encryption</h3><p>Utilize quick-expiring ephemeral tokens to minimize the window of opportunity for attackers in case they get compromised.</p><p>Tokens and sensitive data, in transit or at rest, should always be encrypted to prevent unauthorized interception.</p><h3>5. Log Everything</h3><p>Establish a comprehensive log for all API interactions so that you are always aware of what actions they are performing and when. These logs are crucial for security auditing, incident reporting, and forensic analysis.</p><h3>6. Review Permission Regularly</h3><p>To reduce the attack surface, periodically audit and review all API access permissions to identify and revoke any stale, unused, or excessive access rights. This helps to maintain a strong security posture.</p><h3>IAM Trends to Watch out for</h3><p>With cyber threats getting smarter and users expecting smoother, more secure digital experiences, companies need to stay a step ahead. That means keeping an eye on emerging trends that are reshaping how we manage identity and access. Here are a few you should have on your radar:</p><h3>● Zero Trust Architecture (ZTA)</h3><p>It operates by the <strong><em>‘never trust, always verify’</em></strong> principle, where no user, device, or system is automatically trusted regardless of location. For APIs, every call, whether from within or outside the network, is rigorously authenticated before granting authorization.</p><p>ZTA demands granular access rules that define exactly what each user or service can access and under what conditions.</p><h3>● AI-Driven IAM</h3><p>Artificial Intelligence and Machine Learning are revolutionizing API IAM by enabling proactive and adaptive security. Integrating with AI can help you spot unusual patterns and anomalies that might indicate a security breach and prevent such incidents in real-time.</p><p>For instance, if an API call originates from Tokyo and is immediately followed by one from Boston within three minutes, AI will detect this and flag it as highly suspicious if it is significant deviation from baseline.</p><h3>● Decentralized Identity (DID) &amp; Blockchain</h3><p>It allows secure data sharing without intermediaries. Unlike traditional centralized identities, they use verifiable credentials stored on a blockchain, ensuring that Identity info remains tamper-proof, thus significantly reducing data breaches.</p><p>This approach enhances user privacy and simplifies the verification processes of how identities interact with APIs, especially in sensitive areas like healthcare and finance.</p><h3>Conclusion</h3><p>APIs are the lifeblood of modern digital infrastructure, enabling enterprises to seamlessly share data and resources with other companies, developers, and consumers worldwide.</p><p>However, this very widespread adoption, while enabling unprecedented connectivity and innovation, simultaneously positions APIs as prime targets for cyberattacks. Therefore, robust Identity and Access Management isn’t just a best practice for APIs — it’s an absolute critical imperative.</p><p>Organizations must prioritize continuous monitoring and adaptive access controls to stay ahead of evolving threats. Embracing emerging IAM trends like zero trust and decentralized identity will be key to securing the future of API ecosystems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7852029abfb6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/best-practices-for-api-security-7852029abfb6">Best Practices for API Security</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building the Connected Enterprise: API-First and Cloud-Native Integration Patterns]]></title>
            <link>https://medium.com/@anantrw/building-the-connected-enterprise-api-first-and-cloud-native-integration-patterns-19c64ba51536?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/19c64ba51536</guid>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 03:53:46 GMT</pubDate>
            <atom:updated>2025-11-02T13:13:44.567Z</atom:updated>
            <content:encoded><![CDATA[<h3>Unlock the true power of your systems and data by stitching them effectively.</h3><p>The enterprise environment contains countless data resources, applications, and systems. Ensuring they all work in harmony like a well-oiled machine is vital to the success of any forward-thinking business.</p><p>And while earlier communications between systems were relatively straightforward, today the waters are a bit murkier.</p><p>Almost every tool, app, or system runs on web-based platforms, and the databases are now more sophisticated than ever. What’s more, you still have to contend with customer expectations and competitors.</p><p>So, what’s the smart move here? Simple: unlock the true power of your systems and data by stitching them together properly. That’s where enterprise integration steps in. Let’s break it down.</p><h3>The Cloud Shift — What Changed?</h3><p>Recent data indicate that more than 94% of companies with over 1,000 employees have transitioned most of their work to virtual environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*9R79SrqOwSb3sFEoiDwmUw.png" /><figcaption>Interconnected IT Services</figcaption></figure><p>But that’s just not it, is it? They leverage the best of both worlds as they migrate their workloads to online servers where applications and their components lie in multiple servers, online and on-premises systems.</p><p>This blend offers flexibility, cuts unnecessary costs, reduces dependence on IT staff, and allows easy scalability, just to mention a few. Hence, the shift from legacy systems.</p><p>But this setup isn’t perfect. When you think about connecting these two worlds, you immediately encounter some interesting hurdles. Getting a few plug-ins here and there won’t get you any closer to ensuring these systems work seamlessly with each other.</p><p>Why? Because these systems, whether on-premise, cloud, or hybrid, have entirely different architectures, from security protocols and data formats to how they are designed to operate.</p><p>Getting them to interact effortlessly isn’t easy. It requires a well-thought-out integration strategy focusing on technical aspects and business goals.</p><h3>Key Approaches to Cloud Integration</h3><p>Speaking of integrations, there’s no one-size-fits-all solution. With that in mind, let’s look at three common techniques today.</p><h3>API-First Approach</h3><p>When we say API-first, we mean placing the API at the centre of the development process. Think of it this way: instead of creating applications and then figuring out how they’ll interact with each other, you first prioritize designing reusable APIs.</p><p>With this, you can easily add new services and technologies without re-architecting the system.</p><p>Take Google’s Maps API for example. The API was first designed and well documented, then the code was later developed to execute it. This has allowed Google to make multiple modifications to the API without altering the code that uses it.</p><h3>Cloud-Native Integrations</h3><p>This refers to building, deploying, and managing applications in virtual computing environments. How? By using orchestrators such as Kubernetes, applications are deployed in the form of containers consisting of different microservices, which can be scaled and managed independently.</p><p>Ever wondered how modern applications manage to stay so fast and reliable? This is how. Some real-life examples include: Spotify, which delivers music using cloud-native technology, and Netflix, which seamlessly streams content using microservices.</p><h3>Enterprise Integration Platforms (EiPaaS)</h3><p>This is an integration platform that sits on top of all your systems and applications, helping you to better manage and integrate them. It is like a conductor that orchestrates the flow of data between different systems, thus making it much easier to access and use the data that you need.</p><p>With EiPaaS, you can automate tasks, reduce errors, and improve collaboration across your teams. So, whether you are looking to integrate your CRM with your sales platform or connect a payment system with your procurement system, EiPaaS is your go-to platform.</p><h3>Core Features of Modern Integration Solutions</h3><p>Before we start looking at tools that are making these integrations possible, it’s essential to understand the driving forces behind this process.</p><h3>Data Streaming</h3><p>This is what’s behind real-time analytics. How Netflix is able to predict what you want to watch next, or how enterprises are able to predict future trends.</p><p>IoT devices, web apps, and financial transactions are constantly generating data, which is ingested by streaming platforms like Kafka. This data is analyzed and processed in real time, and then stored in databases or triggers immediate actions like fraud detection.</p><h3>Data Integration</h3><p>So, you’ve got information all over the place, right? It may be hundreds of thousands of rows, if not millions. Chances are, you don’t need every single piece of data from that table.</p><p>Data integration helps us piece it altogether and make sense of it all by:</p><p>● <strong>Data Cleaning</strong>: identifying, correcting (or removing) errors, and removing inconsistencies in data to improve its quality.</p><p>● <strong>Data Transformation</strong>: modifying data to fit a particular purpose or context for better understanding and analysis.</p><h3>Workflow and Process Automation</h3><p>No one likes doing the same tasks over and over. It’s time-consuming, repetitive, and boring. Instead of assigning such tasks to your staff, automation tools can be integrated to perform them at a much faster and accurate rate.</p><p>This is most effective in automating complex multi-processes that involve a lot of approvals, decisions, and steps, especially in cross-departmental teamwork.</p><h3>Security and Governance</h3><p>Integrations by design make it possible for data to flow between different platforms and applications within and outside the enterprise. Unchecked, these integrations could easily expose vital data, resulting in privacy violations, reputational damage, and operational risk.</p><p>To be safe, you need to play by the rules. This means putting in place strong security measures and policies that ensure you lock things down, stay legal, and track how your data is being used.</p><h3>Tools Driving Cloud-Based Integrations</h3><h3>a. Cloudera Data Platform</h3><p>Right off the bat, we have Cloudera Data Platform (CDP), your hybrid data powerhouse. A tool that will help your enterprise turn data into insights in no time.</p><p>And it’s not just one tool, but rather a suite of services designed to work in unison to manage the entire data lifecycle.</p><p>What are some of its core capabilities?</p><h4>Unified Data Lake</h4><p>CDP lets analysts and data scientists access trusted data stored across your hybrid environment so that they can explore on their own with their choice of tools. This enables them to uncover insights exactly when they need them.</p><h4>Data Engineering</h4><p>CDP comes equipped with a wide range of tools for data ingestion, processing, and quality management. These include Apache NiFi, Spark, and Hive.</p><h4>Real-time Analytics</h4><p>Apache Kafka and Flick, components of Cloudera Data Platform, allow your analysts to process and analyze streaming data for real-time insights.</p><h3>b. Cloudera Data Flow</h3><p>As you might have guessed, Cloudera Data Flow exists within the Cloudera ecosystem. Its primary role is to facilitate the rapid development and deployment of data pipelines used to ingest, process, and distribute data from and to any destination.</p><p>How is this done? It’s quite simple. Instead of developers hopping from one extraction tool to another, writing complex code to format and transfer data as you need it, they can instead build and deploy data flows through a quick drag-and-drop interface and manage all of them from one place.</p><h3>c. API Platforms</h3><p>API platforms like Apigee and MuleSoft are used to provide comprehensive solutions for managing the entire lifecycle of Application Programming Interfaces (APIs).</p><p>In this context, they act as the gatekeepers of connectivity, enabling different applications and services to communicate with each other securely regardless of the environment.</p><h4>MuleSoft</h4><p>This is an AnyPoint platform that sits at the intersection of nearly all the biggest technology forces in the world today. It does this by unlocking data using modern APIs and then providing connectivity between systems, external applications, and devices.</p><p>It also enables you to manage and secure the flow of data between all systems in the enterprise, helping create your application network. This allows your team to discover and reuse existing APIs.</p><h4>Apigee</h4><p>Now part of Google Cloud, Apigee is known for its strong API management capabilities, especially in areas such as security, traffic management, and analytics.</p><p>It’s a robust and scalable platform used for exposing and managing APIs at enterprise scale.</p><h3>Conclusion</h3><p>Connecting systems is still a huge challenge for enterprises today. It only seems to get worse as many enterprises scale and shift their workloads to virtual environments.</p><p>Your enterprise could be struggling with disconnected systems, data silos, or inefficient workflows. But this doesn’t need to be the case. You can ensure your systems interact effectively no matter the environment through the proper integrations we’ve covered, and enjoy the benefits of doing so.</p><h3>Resources</h3><p>● <a href="https://celestialsys.com/blogs/modern-enterprise-applications-development-in-the-cloud-era/">Modern Enterprise Applications Development in the Cloud Era</a></p><p>● <a href="https://intverse.medium.com/the-power-of-enterprise-integration-platforms-fueling-organizational-success-dc369b5afcd8">The Power of Enterprise Integration Platforms: Fueling Organizational Success</a></p><p>● <a href="https://boomi.com/blog/what-is-enterprise-platform/">What Is an Enterprise Platform?</a></p><p>● <a href="https://payodatechnologyinc.medium.com/data-integration-in-the-cloud-challenges-and-solutions-aa85b09967cc">Data Integration in the Cloud: Challenges and Solutions</a></p><p>● <a href="https://medium.com/@chirag.dave/what-is-system-integration-a-complete-guide-for-modern-businesses-ec18fba88e58">What is Sytem Intergration? A Complete Guide for Modern Business</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19c64ba51536" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Comparative Assessment of Built-In Security of LLM Models]]></title>
            <link>https://medium.com/design-bootcamp/a-comparative-assessment-of-built-in-security-of-llm-models-1857444c76cb?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1857444c76cb</guid>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Mon, 30 Jun 2025 01:43:17 GMT</pubDate>
            <atom:updated>2025-06-30T11:24:08.903Z</atom:updated>
            <content:encoded><![CDATA[<p>This article establishes a comparative evaluation of built-in security features in prominent LLM models, such as DeepAI’s GPT-4 and Google’s PaLM, together with Cohere’s command model.</p><p>Large language models (LLMs) entering diverse industries require immediate security evaluation since they attract substantial security concerns. These models’ powerful text generation capabilities face security risks involving privacy breaches, data exposure incidents, and targeted attack vulnerabilities (DeepAI, 2025a). Different security threats emerge during LLM implementation and need proper monitoring. The disclosure of sensitive data can occur if LLM training involves confidential information or the system processes this material without appropriate security measures (DeepAI, 2025b). Adversarial attacks can subvert LLM responses when they send deceptive input, which damages the model’s consistent delivery of solutions (Google, 2023). Models experience prompt injection attacks because operators generate particular prompts that produce dangerous responses. Security problems deepen because of ethical content generation difficulties, including spreading misinformation and bias in the systems (DeepAI, 2025b).</p><p><strong>Comparative Assessment of Security Controls</strong></p><p><em>1. DeepAI’s GPT-4</em></p><p>The popular GPT-4 model from OpenAI functions flawlessly due to its potent defense mechanisms. The system automatically includes defensive mechanisms to prevent the creation of damaging or incorrect information during regular operations (DeepAI, 2025a). The platform DeepAI uses reinforcement learning from human feedback (RLHF) to update the model to enhance safety functions within its generated responses. The system employs content-blocking technology, which stops the creation of inappropriate or sexually explicit textual content.</p><p>GPT-4 operates with API access control security through which organizations define their rules for permitted content types and user interaction limits. The technical platform of DeepAI delivers thorough usage policies and guidelines that focus on keeping technology use in line with social values (DeepAI, 2025a). Security concerns exist about possible wrongheaded or biased output from GPT-4 because of its training data.</p><p><em>2. Google’s PaLM</em></p><p>Google advances LLM security significantly by introducing its Pathways Language Model (PaLM). The Pathways Language Model (PaLM) incorporates a system for content moderation that removes dangerous content from its product (Google, 2023). The risk mitigation systems implemented by Google enable users to report issues in generated outputs, which helps develop an interactive feedback process that improves future model performance.</p><p>One notable aspect of PaLM is that it explains how the software derives specific outputs. The approach helps organizations understand potential biases and risks stemming from the model (LeewayHertz, 2023). Google is committed to privacy protection by restricting data retrieval duration with complete user control over its data exploitation and management capability.</p><p>Google has successfully enhanced its security strategies, yet the challenge persists in preserving the consistency of outputs across various situations and protecting against adversarial manipulation (SiliconANGLE, 2023).</p><p><em>3. Cohere’s Command Model</em></p><p>The Command model from Cohere takes a unique approach through its user data security procedures. Cohere integrates encryption tools that secure user data at every entry and exit point of its interaction system by default. Users of the model can control access by roles to minimize unauthorized interaction according to business requirements through built-in access controls.</p><p>Safety features within Cohere enable content generation that satisfies organizational content policies, allowing businesses to adapt the model based on their specific security needs. Like other LLMs this model integrates security features but continues encountering issues about prejudice and potential bad uses.</p><p><strong>Expanding the Security Framework: Effectiveness and Limitations</strong></p><p>The evaluation of security controls in different large language models (LLMs) must examine the measures’ effectiveness and identify their current weaknesses and potential upgrading opportunities within security architectures. Various assessments focusing on LLM technology’s newness depend primarily on research and field experiences developed since each model’s release (Google, 2023; DeepAI, 2025a). We will analyze GPT-4 alongside PaLM and Cohere Command for their security aspects before offering security improvement strategies for LLMs.</p><p><strong><em>Security Effectiveness</em></strong></p><p><em>DeepAI’s GPT-4</em></p><p>GPT-4 builds its security foundation from extensive user collaboration and performance evaluation. Reinforcement learning that uses human feedback stands out as an excellent method to match model responses with human principles (DeepAI, 2025a). Through its advanced capabilities, GPT-4 reduces dangerous content generation that would produce misinformation. Organization staff benefit from the DeepAI API controls because these tools enable them to design experiences that adapt to their organization’s specific security requirements.</p><p>GPT-4 continues to deal with multiple obstacles throughout its operations. The principal drawback pertains to the training data basis since it handles information that might show age or bias effects (DeepAI, 2025b). The model has been designed to reduce biased content generation but suggests that the complete elimination of biased output remains a possibility. Users with prompt engineering expertise can perform adversarial attacks/test cases that lead to model violation of content policies. The necessity for constant security improvement measures becomes essential because threats within the landscape continue to evolve.</p><p><em>Google’s PaLM</em></p><p>Google’s PaLM implements one of the most advanced system safety mechanisms through its content moderation tools combined with transparency functions (Google, 2023). The ability of users to flag outputs along with their feedback enables real-time adjustment logic at the model level that allows it to learn from mistakes. Users can examine model outputs with justification to detect misinformation because of Google’s system for transparency (LeewayHertz, 2023). This essential check will enable users to evaluate model responses.</p><p>These feedback systems used in PaLM demonstrate promising capabilities, yet they create susceptibility points for the model. User-generated improvement inputs depend on sufficient user participation because insufficient user flags for problematic output can result in unaddressed serious issues (SiliconANGLE, 2023). The benefit of transparency in AI systems becomes limited because complex explanations about AI operations remain understandable to those who lack expertise in this field.</p><p><em>Cohere’s Command Model</em></p><p>The models implemented by Cohere for protecting data privacy through encryption methods combined with access control measures provide exceptional benefits to organizations that must follow strict data management protocols. The model protects interaction data effectively, which leads to a substantial reduction in potential information loss risks. Cohere addresses the misuse potential of its technology by enabling organizations to customize the content generation process according to their internal policies.</p><p>Cohere faces limitations similar to other LLMs because it encounters challenges in defending against prompt injection attacks and unintentional bias within its system. Although security measures stop specific types of abuse from happening, adversaries adapt their methods to find new undefended weaknesses (DeepAI, 2025b). To maintain security with Cohere’s technology, companies have to perform frequent evaluations of these safety measures alongside the development of mechanisms for addressing newly discovered threats.</p><p><strong>Addressing Limitations: Strategies for Improvement</strong></p><p>The durability of security depends on proper solutions for the architectural weaknesses discovered in the studied LLM systems. Multiple measurable strategies follow to improve the performance as mentioned below:</p><ol><li><strong><em>Continuous Learning and Adaptation:</em></strong> A continuous feedback system enables real-time human interactions to improve the responses generated by LLMs. These systems acquire their evolution capabilities through the insights of multiple users, particularly experts who enable navigation through changing contexts and user demands alongside threat discovery (DeepAI, 2025a).</li><li><strong><em>Fine-tuning and Retraining:</em></strong> Organizations must plan systematic model retraining as part of their routine operations because new data situations and environments emerge. The integration of modern dataset collections that match current social standards makes outdated or biased model outputs less likely, so the model becomes more reliable (Google, 2023).</li><li><strong><em>Adversarial Training:</em></strong> Organizations should fight adversarial attacks by training their models to encounter dangerous prompts during their development process. The security framework training process exposes the models to malicious attack approaches to detect such attempts before they compromise system security (DeepAI, 2025b).</li><li><strong><em>Enhanced Transparency and User Training:</em></strong> Making LLM operations transparent demands parallel user education and awareness programs for both programmers and end-users about LLM functionalities and capabilities. The approach allows users to use models effectively while ensuring security,, resulting in maximum operational potential (LeewayHertz, 2023).</li><li><strong><em>Collaboration for Ethical Standards:</em></strong> LLM developers must work together with regulatory bodies, academic institutions, and ethical oversight organizations to establish preferred practices in safe model deployment. Organizations should place ethical weight on applications,, particularly in healthcare, financial services, and legal practices, because severe consequences may occur (SiliconANGLE, 2023).</li><li><strong><em>Monitoring and Assessment Protocols:</em></strong> Model maintenance through regular output audits enables the identification of new security risks and their timely resolution. Organizations must create extensive monitoring systems that will allow them to identify and solve potential risks throughout their operations (Google, 2023).</li></ol><p><strong>Conclusion: Forward Path for LLM Security</strong></p><p>The expanding integration of DeepAI’s GPT-4, Google’s PaLM, and Cohere’s Command requires priority security measures. Each model brings distinct advantages, which include advanced assessments for contents alongside transparent explanations about results and robust data security systems. Continuous evaluation and innovation become necessary because LLM security faces powerful weaknesses stemming from adversarial factors and restricted feedback functionalities (DeepAI, 2025a; Google, 2023).</p><p>The modern technological evolution requires security approaches that evolve dynamically through multiple evaluation methods to manage built-in security threats and execute proactive solutions adequately.</p><p>Organizations should engage stakeholders through dialogue as they provide education programs about advanced models to establish responsible usage practices for LLM technology (LeewayHertz, 2023). The long-term objective involves expanding our comprehension of AI dangers to establish broad LLM usage while reducing possible risks. Selecting LLMs should focus on finding systems that maintain strong security measures while maintaining their innovative capabilities (SiliconANGLE, 2023).</p><p><strong>References</strong></p><p>DeepAI. (2025). <em>GPT-4 technical report</em>. DeepAI. Retrieved from <a href="https://deepai.org/publication/gpt-4-technical-report">https://deepai.org/publication/gpt-4-technical-report</a></p><p>DeepAI. (2025). <em>Unveiling security, privacy, and ethical concerns of ChatGPT</em>. DeepAI. Retrieved from <a href="https://deepai.org/publication/unveiling-security-privacy-and-ethical-concerns-of-chatgpt">https://deepai.org/publication/unveiling-security-privacy-and-ethical-concerns-of-chatgpt</a></p><p>SiliconANGLE. (2023, April 24). <em>Google Cloud bolsters cybersecurity with generative AI model Sec-PaLM</em>. SiliconANGLE. Retrieved from <a href="https://siliconangle.com/2023/04/24/google-cloud-bolsters-cybersecurity-generative-ai-model-sec-palm/">https://siliconangle.com/2023/04/24/google-cloud-bolsters-cybersecurity-generative-ai-model-sec-palm/</a></p><p>Google. (2023, May 10). <em>Google AI: What to know about the PaLM 2 large language model</em>. Google. Retrieved from <a href="https://blog.google/technology/ai/google-palm-2-ai-large-language-model/">https://blog.google/technology/ai/google-palm-2-ai-large-language-model/</a></p><p>LeewayHertz. (2023, May 15). <em>Google PaLM 2: Features, applications and development</em>. LeewayHertz. Retrieved from <a href="https://www.leewayhertz.com/google-palm2/">https://www.leewayhertz.com/google-palm2/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1857444c76cb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/a-comparative-assessment-of-built-in-security-of-llm-models-1857444c76cb">A Comparative Assessment of Built-In Security of LLM Models</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Evolution of API Design: A Systematic Evaluation of Six Key Architectures Powering Modern…]]></title>
            <link>https://medium.com/design-bootcamp/the-evolution-of-api-design-a-systematic-evaluation-of-six-key-architectures-powering-modern-5846b6402aa2?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/5846b6402aa2</guid>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Fri, 27 Jun 2025 04:36:27 GMT</pubDate>
            <atom:updated>2025-06-27T11:08:57.488Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Evolution of API Design: A Systematic Evaluation of Six Key Architectures Powering Modern Digital Ecosystems</h3><p>The rising level of global connection requires powerful APIs to function effectively. Current applications depend on APIs to facilitate communication between different computing services and platforms. Multiple architectural approaches emerged to support these communications between systems with their own characteristics for specific use requirements. This research evaluates six widely used API architecture methods: Restful, GraphQL, SOAP, gRPC, WebSockets, and MQTT.</p><p><strong>1. Restful APIs</strong></p><p>The representational state transfer (REST) architectural style has become popular because it delivers both a simple design and efficient outcomes. RESTful APIs enable communication through standard HTTP commands, including GET, POST, PUT, and DELETE (IBM, 2025). A key requirement of REST is statelessness, which demands that every client request provide complete data needed by the server for processing (Fielding, 2000). The server operating in this stateless condition enables scalability because it avoids keeping track of session information between requests.</p><p>The principles that form the base of REST also encompass cache-ability and a uniform interface. Implementing cache-ability permits servers to save responses for later reuse, decreasing request frequency and response time (IBM, 2025). The interface standardization facilitates component code interoperability so different services can easily connect. REST remains independent from the data format used for exchange through JSON, although it commonly pairs with this format.</p><p>REST provides a popular platform for web services because of its ease of use. Still, its popularity is limited by its inability to meet complex querying or real-time data retrieval requirements (IBM, 2025).</p><p><strong>2. GraphQL</strong></p><p>The Facebook-developed GraphQL system replaces REST interfaces with a mechanism that let clients precisely choose their required data sets. This functionality stops clients from obtaining too much data while preventing them from receiving inadequate amounts of data during a single request, enhancing the data retrieval process (GraphQL Foundation, 2025). GraphQL enables clients to request the server by providing information about data specifications and required fields that the server uses to produce its response.</p><p>The data requirements of application systems with complex interdependent resources provide the perfect environment for GraphQL to produce benefits. Through GraphQL, developers can operate with one endpoint, which decreases the complexity of integration while lowering required data collection request numbers (Solo.io, 2025). The subscription feature of GraphQL provides real-time data updates, which work well for applications that need speed in information synchronization (GraphQL Foundation, 2025).</p><p>The ability of GraphQL to handle complex server implementations leads to increased complexity because developers need to establish schemas and resolvers for appropriate query processing. Implementing GraphQL requires thorough examination and extensive planning to receive all its advantages despite the existing risk factors (Solo.io, 2025).</p><p><strong>3. SOAP</strong></p><p>Simple Object Access Protocol (SOAP) is a standard protocol that enables application communication through XML message exchanges. SOAP APIs implements rigid specifications and data structure design, allowing their successful application in enterprise systems with strong transaction security and integrity requirements (Box et al., 2000). SOAP provides complex business transaction management and secure communication through WS-Security protocol implementation.</p><p>A SOAP contract is defined through the Web Services Description Language (WSDL), which prescribes service methods, message formats, and available operations (Box et al., 2000). SOAP’s self-defining aspect creates structured applications that help developers understand and build services more efficiently.</p><p>SOAP’s complex and rigid nature leads to challenges during new project implementation, primarily when projects must compete with the more straightforward REST and GraphQL APIs. SOAP continues to serve applicable industries requiring standards and compliance, including finance and healthcare operations (Box et al., 2000).</p><p><strong>4. gRPC</strong></p><p>Through its gRPC (gRPC Remote Procedure Call) framework, Google uses Protocol Buffers to achieve efficient data serialization, which results in fast performance and minimal response time in communication (Google, 2025). gRPC serves as a purpose-built solution for microservices communication inside service-oriented systems because its operation focuses on maximum efficiency within quick response time needs.</p><p>GRPC serves various programming languages, creating versatility for developers in different programming environments. The protocol applies HTTP/2 technology that lets users enjoy multiplexing, header compression, and bi-directional streaming for better real-time data communication (Google, 2025). gRPC works best for mobile and distributed systems because of its efficient connection abilities.</p><p>Adopting Protocol Buffers allows developers to learn a distinctive data serialization method that brings initial onboard costs. GPRC lacks complete API browser support similar to REST and GraphQL, rendering it ineffective for some front-end system applications (Google, 2025).</p><p><strong>5. WebSockets</strong></p><p>WebSockets enable servers and clients to build two-way live connections through a standardized communication protocol. Unlike standard HTTP stateless requests, an open WebSocket connection provides continuous data transmission (Fette &amp; Melnikov, 2011). WebSockets offer an optimal solution for applications needing instant data updates in platforms like chat applications, gaming platforms, and collaborative tools.</p><p>WebSockets establish single, long-lasting bidirectional connections, which reduces resource consumption when repeatedly setting up and closing connections. WebSockets excel in applications that need minimal latency as their operating efficiency provides significant value to time-sensitive scenarios (Fette &amp; Melnikov, 2011).</p><p>Developing WebSockets applications requires skilled connection control methods and state administration, leading to architectural complexity. Additional backup systems will be needed when WebSockets do not receive support (Fette &amp; Melnikov, 2011).</p><p><strong>6. MQTT</strong></p><p>The Message Queuing Telemetry Transport protocol (MQTT) is a lightweight messaging protocol explicitly made for low bandwidth and high latency situations. Before its adoption in IoT applications, MQTT served remote monitoring, and control purposes and continues to thrive as a standard solution for IoT devices that need efficient communication methods (OASIS, 2019).</p><p>MQTT utilizes a publish-subscribe method that enables producers and consumers to remain separate from each other before messaging each other. The system’s structure enables adaptable workflow patterns and simple expansion capabilities, which fit nicely in applications requiring simultaneous device interaction (OASIS, 2019).</p><p>One exceptional feature of MQTT is its fitting application to IoT implementations, but other applications need sophisticated queries or high levels of security beyond its capabilities. The selection of MQTT as a communication protocol should follow application requirements when developing new software systems (OASIS, 2019).</p><p><strong>Conclusion</strong></p><p>The choice of appropriate API architecture style represents a key element to producing applications that have high efficiency, scalable characteristics, and ensure maintainability. Strong and weak aspects of RESTful, GraphQL, SOAP, gRPC, WebSockets, and MQTT allow these protocols to support various operational scenarios. Understanding architectural characteristics and application domains will enable developers to pick styles that perfectly fit their project requirements, thus improving user experiences and better application performance.</p><p><strong>References</strong></p><p>Box, D., Ehnebuske, D., Kakivaya, G., Layman, A., Mendelsohn, N., Nielsen, H. F., … &amp; Winer, D. (2000). <em>Simple object access protocol (SOAP) 1.1</em>. W3C. <a href="https://www.w3.org/TR/SOAP/">https://www.w3.org/TR/SOAP/</a></p><p>Fette, I., &amp; Melnikov, A. (2011). <em>The WebSocket protocol</em> (RFC 6455). IETF. <a href="https://datatracker.ietf.org/doc/html/rfc6455">https://datatracker.ietf.org/doc/html/rfc6455</a></p><p>Fielding, R. T. (2000). <em>Architectural styles and the design of network-based software architectures</em> (Doctoral dissertation, University of California, Irvine).</p><p>Google. (2025). <em>Introduction to gRPC</em>. <a href="https://grpc.io/docs/what-is-grpc/introduction/">https://grpc.io/docs/what-is-grpc/introduction/</a></p><p>GraphQL Foundation. (2025). <em>GraphQL best practices</em>. <a href="https://graphql.org/learn/best-practices/">https://graphql.org/learn/best-practices/</a></p><p>IBM. (2025). <em>What is a REST API?</em> <a href="https://www.ibm.com/think/topics/rest-apis">https://www.ibm.com/think/topics/rest-apis</a></p><p>OASIS. (2019). <em>MQTT version 5.0</em>. <a href="https://docs.oasis-open.org/mqtt/mqtt/v5.0/mqtt-v5.0.html">https://docs.oasis-open.org/mqtt/mqtt/v5.0/mqtt-v5.0.html</a></p><p>Solo.io. (2025). <em>GraphQL: How it works, examples, and 4 best practices</em>. <a href="https://www.solo.io/topics/graphql">https://www.solo.io/topics/graphql</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5846b6402aa2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/the-evolution-of-api-design-a-systematic-evaluation-of-six-key-architectures-powering-modern-5846b6402aa2">The Evolution of API Design: A Systematic Evaluation of Six Key Architectures Powering Modern…</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Architecting Intelligence: Enterprise Design Patterns for GEN AI]]></title>
            <link>https://medium.com/design-bootcamp/architecting-intelligence-enterprise-design-patterns-for-gen-ai-813a7db99c0c?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/813a7db99c0c</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[design]]></category>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Wed, 05 Mar 2025 05:23:17 GMT</pubDate>
            <atom:updated>2025-03-05T14:03:31.090Z</atom:updated>
            <content:encoded><![CDATA[<p>This article presents the most effective enterprise design patterns which enable smooth development and deployment of generative AI applications.</p><p>The implementation of generative AI systems changed traditional methods of performing basic tasks which people used during their daily routines. This technology enables users to obtain fast answers and generate excellent content while reaching higher productivity goals during shorter periods. The technological advancement has reshaped both business management and customer-business partnership mechanics. The combination of AI-based applications along with GEN AI functionality in current programs allows organizations to deliver better service to consumers through enhanced accuracy alongside superior efficiency.</p><h3>Enterprise Design Patterns for Generative AI Applications</h3><h3>Key Features of a Generative AI Application</h3><p>The following segment explores the fundamental elements that enable an efficient GenAI system after exploring its essential components:</p><p>· <strong>Large Language Models: </strong>Large Language Models serve as the backbone of any GenAI system because they hold the most essential role. Large language models receive extensive data training which makes them produce outputs that mimic human writing and logical sense.</p><p>· <strong>Prompt Engineering:</strong> Prompt engineering refers to providing the LLMs with the prompts that are able to define the user needs to the model.</p><p>· <strong>Retrieval Augmented Generation:</strong> This feature enhances the model’s efficiency by retrieving additional information from external sources.</p><p>· <strong>Embeddings:</strong> Used for semantic search, these are numeric representations of text and other data.</p><p>· <strong>Evals or Evaluations:</strong> Used for improving the model, the evaluations evaluate the accuracy and efficiency of the model’s responses.</p><h3>What Are Enterprise Design Patterns?</h3><p>A collection of standard solutions exists to address common problems in design patterns. The enterprise design patterns have the same status as the other patterns. They are the collection of solutions to the problems that occur while enterprise software is being developed. They are highly crucial for robust, accurate and complex system development. They boost the development speed and code quality and enhance code maintainability, especially when developing a complex system like a generative AI application.</p><h3>Enterprise Design Patterns for Generative AI</h3><p>The following are the best enterprise design patterns for developing generative AI applications:</p><h3>1. Prompt Engineering: Driving the AI Narrative</h3><p>Defining and designing the prompts in a way that the model is able to completely understand the user’s desired output is called <strong>prompt engineering</strong>. It includes the prompt creation, prompt versioning and testing. The LLM’s performance is enhanced by providing the model with engineered prompts, and the responses are generated to align with the user’s needs.</p><p>The prompt engineering can be done as follows:</p><p>v Create template prompts that users can reuse by making small changes to ensure consistency and efficiency. These are known as <strong>Template-Based Prompts</strong>.</p><p>v Improve the AI response’s accuracy by providing the relevant context, known as <strong>Context-Based Prompts</strong>.</p><p>v By defining and assigning different roles to AI, the AI’s responses can be personalized according to your situation. This is known as <strong>Role-Based Prompts</strong>.</p><p>v The <strong>Few-Shot or Zero-Shot Learning</strong> can be incorporated with sample outputs (few-shot) or structured prompts (Zero-shot) to help models be productive.</p><h4>Implementing the Prompt Engineering Pattern</h4><p>The implementation of <strong>prompt engineering</strong> involves the following tasks:</p><p>Ø Developing the template prompts that can be used with slight modifications.</p><p>Ø Tracking the changes in AI’s responses by treating the prompts as code with version control.</p><p>Ø With the help of automated or human-centred tests, evaluate the prompt’s performance and accuracy.</p><p>Ø Collect the useful prompts into a centralized library accessible to everyone easily.</p><h3>2. Retrieval Augmented Generation (RAG) Pattern: Augmenting Knowledge</h3><p>The LLMs are generally trained on huge data sets that cover almost all the subjects and topics; however, they lack the most recent knowledge. Therefore, Retrieval Augmented Generation (RAG) is incorporated to help LLM find such information by accessing external data sources. It fetches the relevant information and documents from the external knowledge bases. It has two parts:</p><p><strong>i.</strong> <strong>Retriever: </strong>Responsible for searching and fetching the documents.</p><p><strong>ii.</strong> <strong>Generator: </strong>Responsible for transforming the retrieved data into a usable response.</p><p><strong>RAG</strong> involves:</p><p>v Fetching information from external knowledge bases and databases.</p><p>v Representing the semantic meaning of text numerically (embeddings) to help LLMs differentiate different datasets.</p><h4>Implementing Retrieval Augmented Generation (RAG)</h4><p>It involves the following tasks:</p><p>Ø Indexing the relevant documents and data sources with the help of <strong>embeddings</strong>.</p><p>Ø Using embeddings to find the most relevant information according to the prompt.</p><p>Ø Converting the fetched data into useful responses.</p><h3>3. Orchestration Pattern: Building Robust Systems</h3><p>The orchestration pattern implies breaking down the GenAI applications into independent microservices to enhance the flexibility and scalability of the system. It helps better manage collaboration between different microservices.</p><p>It involves:</p><p>v Enhancing the model’s efficiency by saving the frequently used prompts and data (known as <strong>Caching</strong>).</p><p>v Improving the model’s speed with the help of GPUs.</p><p>v Clearly defining the flow of data between services and different system components.</p><p>v The system requires real-time monitoring through prompt logging to maintain its performance and detect as well as repair malfunctions.</p><h4>Implementing Orchestration Model</h4><p>The execution of the orchestration model requires completion of several tasks listed below:</p><p>Ø The application requires decomposition into smaller independent operational units also known as microservices.</p><p>Ø The system will receive incoming requests through APIs which transmit the data to proper functional services.</p><p>Ø The longer execution tasks should go through message queues for asynchronous processing implementation.</p><p>Ø The microservice implementation requires orchestration tool Kubernetes for workflow management and deployment..</p><h3>4. Evaluation Pattern: Ensuring Quality and Reliability</h3><p>Through this pattern you can keep tabs on how accurate and efficient as well as how reliable the AI-generated responses remain. The pattern enables you to evaluate both qualitative and quantitative metrics including:</p><p>- <strong>Accuracy and Relevance: </strong>Are the responses factually correct?</p><p>- <strong>Coherence and Fluency: </strong>Are the responses well-structured?</p><p>- <strong>Toxicity and Bias Detection: </strong>Are the responses compliant with ethical guidelines?</p><p>The evaluation pattern involves:</p><h4>Implementing Evals Pattern</h4><p>The implementation of evaluation pattern requires execution through these steps:</p><p>Ø The model gets assessed through quantitative evaluation of its accuracy and performance metrics.</p><p>Ø A necessary step within this process requires human evaluation to measure model performance.</p><p>Ø The repetitive model evaluation allows teams to find areas that require improvement.</p><p>Ø The model needs proper testing to detect adversarial attacks.</p><p>Ø A model evaluation process requires assessment of its response quality which includes relevance together with coherence and fluency.</p><h3>5. Governance and Responsible AI Pattern: Ensuring Ethical Deployment</h3><p>Within this pattern organizations establish framework principles that determine ethical conduct and governance practices when developing AI applications. To achieve successful deployment of enterprise generative AI the most important elements are security alongside ethical governance practices. It involves:</p><p>v The system protects itself from unauthorized intrusions by following all privacy policies and regulations.</p><p>v The implementation of strong access control systems serves to control data access.</p><p>v The model needs to deliver explanations about its behaviors together with its responses and decision-making process.</p><p>v All organizations need to follow ordinary accepted security measures along with established ethical rules.</p><h4>Implementing Security and Governance AI Pattern</h4><p>Ø The elimination process of biases which exist in both data sets and user responses requires proper recognition.</p><p>Ø The system must demonstrate clear explanations about its processes to arrive at particular judgments.</p><p>Ø General data security procedures should be followed for model protection.</p><p>Ø A deployment of ethical policies for responsible AI operations.</p><h3>6. Query Rewriting and Reranking Pattern: Enhancing Search</h3><p>The user sometimes presents an unclear explanation to the model regarding their requirements. Application development in generative AI heavily depends on the coupling of query rewriting and reranking functions. Problems arise when complex queries are sent directly to retrieval systems because suboptimal results occur so query rewriting patterns provide models with user need comprehension. The process of query rewriting involves both the optimization of user requests along with the elimination of confusing and repetitious information and the process of query reranking emphasizes boosting the usefulness of retrieved documents and data.</p><h4>Query Rewriting Techniques</h4><p>Ø <strong>Finding Synonyms: </strong>The technique involves replacing original words with their synonym lists to increase document relevancy.</p><p>Ø <strong>Grammar Correction: </strong>Correcting any grammar mistakes.</p><p>Ø <strong>Recognizing the Intent: </strong>The model needs to interpret the user intent through the provided prompt.</p><h4>Reranking Techniques</h4><p>Ø <strong>BM25 and Neural Reranking: </strong>Incorporating deep-learning-based ranking with traditional retrieval.</p><p>Ø <strong>Context-Aware Reranking: </strong>The system utilizes user history together with preference settings to create result lists.</p><p>Ø <strong>Reinforcement Learning: </strong>Ranking based on user feedback.</p><h3>Conclusion</h3><p>The combination of popularity with usability and effectiveness and accuracy of generative AI systems is driving businesses to implement them in their systems.</p><p>The <strong>prompt engineering</strong> and <strong>retrieval augmented generation (RAG)</strong> are crucial for enhancing LLM’s accuracy and efficiency. <strong>Orchestration</strong> reduces the inter-dependency of microservices. <strong>Evaluation pattern</strong> ensures relevancy, coherence and fluency, <strong>responsible AI</strong> ensures ethical deployment and <strong>rewriting and reranking</strong> guarantees fulfilled user needs. With a strategic approach, organizations can explore the true potential of generative AI capabilities and be more productive.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=813a7db99c0c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/architecting-intelligence-enterprise-design-patterns-for-gen-ai-813a7db99c0c">Architecting Intelligence: Enterprise Design Patterns for GEN AI</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seven Layers of AI Model Architecture: A Deep Dive into the Foundation of AI Systems]]></title>
            <link>https://medium.com/design-bootcamp/seven-layers-of-ai-model-architecture-a-deep-dive-into-the-foundation-of-ai-systems-a8d9511c448a?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8d9511c448a</guid>
            <category><![CDATA[design]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Wed, 05 Mar 2025 03:06:12 GMT</pubDate>
            <atom:updated>2025-03-05T14:03:37.358Z</atom:updated>
            <content:encoded><![CDATA[<p>The article goes on to describe the first four levels of the AI model architecture and the functional interrelationships between them in the functional sense that they form efficient and performance-driven AI programs.</p><p>The proficiency of Artificial Intelligence in different businesses through its modern capacities gives prime help to human guidance. Although these AI systems work smoothly because of their multi-layer core structure, the core of AI systems remains unnoticed.</p><p><strong>1. Physical Layer: The Foundation of AI</strong></p><p>All computational operations in AI architecture are carried out from the physical layer. The first layer of AI model architecture is based on traditional as well as state-of-the-art infrastructure and hardware components. Physical Layer uses GPUs, TPUs, and quantum computing to perform quick computations and data handling operations (Jouppi et al., 2017).</p><p>The bottom layer is irreplaceable because it regulates speed, how fast an AI application runs, and how expandable it is. There are two options for hosting organizations’ AI models: use of cloud services like AWS using Google Cloud Platform and Microsoft Azure or keeping the servers in an on-site facility. By physically layering optimization, businesses can run their AI applications at high workload conditions (Patterson et al., 2021).</p><p>Fundamentally, edge device advancement breaks down artificial intelligence processing into something different. With the rise in IoT devices, the practice for AI processing becomes to do that on edge loca­tions rather than a centralized data facility to solve latency and bandwidth problems. The need for edge computing is a gap since autonomous vehicles need to analyze accurate real-time environment information; therefore, the computing approach is necessary (Jouppi et al., 2017).</p><p><strong>2. Data Link Layer: Bridging AI with Real-World Applications</strong></p><p>Physically extracting the data from infrastructure is the most crucial task of integrating the AI models into the actual operational environment. These functions are provided using the Data Link Layer by model serving and API integration. Through this bridging system, practical user experience (Sculley et al., 2015) is provided for users to engage with functional AI applications that enable AI operational capability.</p><p>This layer deploys models efficiently using MLOps practices and the LangChain AutoGPT orchestration system that helps manage AI operations workflow using deployment management (Abadi et al., 2016). MLOps (Machine Learning Operations) is the view in which the tools operate, in which AI models receive lifecycle management capabilities (including performance and scalability), as well as capabilities for both lifecycle operational growth (Sculley et al., 2015).</p><p>Connecting artificial intelligence systems to an organization’s current system infrastructure is done smoothly with API-driven AI services. However, with the application programming interfaces of Software as a Service provider, they are able to introduce AI features that only enhance their original functionality. This critical role of the Data Link Layer is crucial in the fundamental value of meaningful business from AI — the chatbots and recommendation engines AI uses today (Shum et al., 2018).</p><p><strong>3. Computation Layer: The Brain of Real-Time Processing</strong></p><p>The Computation Layer can provide the link to logical task execution and processing so as to make AI possible. This layer supports high-performance AI applications with the help of advanced accelerators like GPUs, Distributed Computing methods, and TPUs. Through operations happening in the Computation Layer (Dean et al., 2012), Real-time processing obtains its magic status.</p><p>Common AI frameworks such as PyTorch, TensorFlow, and JAX are given by the framework layer and used to provide the tools used in model development and execution (Abadi et al., 2016). Businesses can perform tremendous-scale server-based calculations using distributed computing, enabling firms to process enormous datasets in the determined life cycles and in actual time for financial security and marketing optimization needs (Dean et al., 2012).</p><p>It gives edge artificial intelligence an added layer to run its operations without sending information to centralized servers. This technology makes the reduction of latency and bandwidth usage (Jouppi et al., 2017) possible and thus able to provide immediate feedback and communication in smart home devices and personal assistants.</p><p><strong>4. Knowledge Layer: Enhancing AI Decision-Making</strong></p><p>After the last layer of computations, the layer responsible for enhancing AI models through external data integration improves their decision capabilities (Hogan et al., 2021). Based on Retrieval-Augmented Generation (RAG) with knowledge graphs and vector search technology (Lewis et al., 2020), the Knowledge Layer uses vector search technology.</p><p>Vast information databases strengthen AI’s knowledge layer with the aim of enhancing AI interactions. Computational systems process wide-ranging databases to generate query responses and validate the data as delivery of the appropriate recommendations in distinct situations (Hogan et al., 2021). It can work in both Google search platforms and on advanced developing tools such as GitHub Copilot based on its ability to provide programming suggestions from a programming dataset with large amounts (Chen et al., 2021).</p><p><strong>5. Model Training and Optimization on the Learning Layer</strong></p><p>The central function of AI is the learning layer, which is focused on model training and optimization processes. It fits an established set of algorithms for making predictors, which basically means classifying assessments based on data that is given. Machine learning and deep learning methods such as Neural Networks and Transformers, Decision trees, Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs). (LeCun et al., 2015).</p><p>The optimization methods of backpropagation, gradient descent, and reinforcement learning serve fundamental functions in model performance enhancement through this layer. The multiple techniques demonstrate different methods that guide parameter adjustment to achieve minimum prediction errors. Continual learning is essential for systems that apply facial recognition and self-driving cars because it enables them to adapt smoothly to different environments and data (LeCun et al., 2015).</p><p>AI model training demands significant amounts of large datasets coupled with high-performance computing requirements. Efficient integration with previous layers, starting with the Physical and Computation Layers, is essential because of this requirement (Patterson et al., 2021). Organizations gain the ability to develop more substantial and more precise AI models because they can handle extensive training data efficiently through cloud platforms as well as edge devices.</p><p>AI models must undergo continuous improvement cycles using retraining to guarantee their adaptation to changing data trends. Organizations utilizing persistent feedback integration systems in their Learning Layer gain improved abilities to resolve market changes and consumer demands.</p><p><strong>6. Representation Layer: Transforming Raw Data into Meaningful Insights</strong></p><p>Data quality determines AI model performance as it is processed in the Representation Layer. This processing system gives the raw data that AI models would be able to understand in a form that is understandable to anyone and would give a structure for efficient data operations. The data processing framework is composed of tokenization, vectorization embedding, and normalization (Mikolov et al., 2013).</p><p>Critical processes such as TF-IDF (Term Frequency-inverse Document Frequency), Word2Vec, and BERT embeddings (Devlin et al., 2019; Mikolov et al., 2013) make text data transformation into numerical entities that deliver semantic meaning, a possibility through the use of NLP. In other words, computer vision is rendering images into numerical arrays for algorithmic analysis, which can identify objects, facial attributes, and movements.</p><p>Feature engineering is an objective that demands this transformation in order to obtain the very basis of data that it will feed the model based on with filtered-out data that it does not really need. This representation strategy works well in improving predictive models as well as increasing the processing flow in order to gain right away in terms of chatbot operation, recommendation service, and predictive analytic function.</p><p>As has been seen, whatever complicated algorithm you craft out, if it does not correctly understand the data, then you will fail unless you have an appropriate data representation approach.</p><p><strong>7. Application Layer: The Interface Between AI and Users</strong></p><p>The last step in our journey into layers is the Application Layer, which is the interface through which AI models communicate with users of real-world applications (Shum et al., 2018). The combination of several chatbots, virtual assistants, automation tools, and API as user interfaces lets the users connect with the AI systems (Radford et al., 2019).</p><p>In this layer, the Applications, in turn, turn the capabilities offered by the preceding six layers into practical interfaces that users can easily exploit (Radford et al., 2019). ChatGPT, Bard, and Claude all show three important LLM (Large Language Model) applications in humanlike responses that, in turn, increase user productivity and engagement.</p><p>There are multiple application options in this layer. The AI in the Application Layer is used by the user directly for the purposes of developing articles or marketing content and also for user interaction with AI, offering the user AI-powered customer support.</p><p>The main priority in this development phase is user interface design. By guiding users quickly through interfaces, eliminating complexity from their AI interactions, and offering sophisticated AI features, users can use advanced AI in ways that do not disrupt their understanding and work. Workflow automation tools highlight the Application Layer because they present how AI optimizes business and keeps better choices with a rise in efficiency.</p><p><strong>Conclusion: The Integrated Layers of AI Architecture</strong></p><p>The seven layers of AI Model Architecture are like a single system, and each layer provides its unique capability to build practical versions of AI. For organizations to achieve maximum benefit from the use of AI, it is essential to understand the fundamental physical layer, which should have an application layer oriented toward the users (Vaswani et al., 2017).</p><p>Finally, and of course, companies must study the peculiarities of each layer of AI for maximum results, but built in the right way. Through these methods, companies will be able to build strong AI systems that will result in solutions that satisfy existing needs and can adjust to upcoming challenges, resulting in innovative outcomes in our complex environment. With thorough AI architecture knowledge, you can make such successful technology management, which leads to an organization’s investment success and innovative development of the future.</p><p><strong>References</strong></p><p>Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., … &amp; Zheng, X. (2016). TensorFlow: A system for large-scale machine learning. <em>12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16)</em>, 265–283.</p><p>Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H. P., Kaplan, J., … &amp; Zaremba, W. (2021). Evaluating large language models trained on code. <em>arXiv preprint arXiv:2107.03374</em>.</p><p>Dean, J., Corrado, G., Monga, R., Chen, K., Devin, M., Mao, M., … &amp; Ng, A. Y. (2012). Large scale distributed deep networks. <em>Advances in Neural Information Processing Systems</em>, 25, 1223–1231.</p><p>Devlin, J., Chang, M. W., Lee, K., &amp; Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. <em>Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies</em>, 1, 4171–4186.</p><p>Hogan, A., Blomqvist, E., Cochez, M., d’Amato, C., de Melo, G., Gutiérrez, C., … &amp; Maleshkova, M. (2021). Knowledge graphs. <em>ACM Computing Surveys (CSUR)</em>, 54(4), 1–37.</p><p>Jouppi, N. P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., … &amp; Yoon, D. (2017). In-datacenter performance analysis of a tensor processing unit. <em>Proceedings of the 44th Annual International Symposium on Computer Architecture</em>, 1–12.</p><p>LeCun, Y., Bengio, Y., &amp; Hinton, G. (2015). Deep learning. <em>Nature</em>, 521(7553), 436–444.</p><p>Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … &amp; Riedel, S. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. <em>Advances in Neural Information Processing Systems</em>, 33, 9459–9474.</p><p>Mikolov, T., Chen, K., Corrado, G., &amp; Dean, J. (2013). Efficient estimation of word representations in vector space. <em>arXiv preprint arXiv:1301.3781</em>.</p><p>Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L. M., Rothchild, D., … &amp; Zhou, L. (2021). Carbon emissions and large neural network training. <em>arXiv preprint arXiv:2104.10350</em>.</p><p>Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., &amp; Sutskever, I. (2019). Language models are unsupervised multitask learners. <em>OpenAI Blog</em>, 1(8), 9.</p><p>Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., … &amp; Dennison, D. (2015). Hidden technical debt in machine learning systems. <em>Advances in Neural Information Processing Systems</em>, 28, 2503–2511.</p><p>Shum, H. Y., He, X., &amp; Li, D. (2018). From Eliza to XiaoIce: Challenges and opportunities with social chatbots. <em>Frontiers of Information Technology &amp; Electronic Engineering</em>, 19(1), 10–26.</p><p>Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … &amp; Polosukhin, I. (2017). Attention is all you need. <em>Advances in Neural Information Processing Systems</em>, 30, 5998–6008.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8d9511c448a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/seven-layers-of-ai-model-architecture-a-deep-dive-into-the-foundation-of-ai-systems-a8d9511c448a">Seven Layers of AI Model Architecture: A Deep Dive into the Foundation of AI Systems</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mitigating Bias in AI Systems]]></title>
            <link>https://medium.com/design-bootcamp/mitigating-bias-in-ai-systems-b2894657078c?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/b2894657078c</guid>
            <category><![CDATA[design]]></category>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Mon, 03 Mar 2025 05:21:22 GMT</pubDate>
            <atom:updated>2025-03-06T03:57:00.154Z</atom:updated>
            <content:encoded><![CDATA[<p>Automatic Systems are ubiquitous today: one uses them to look for a job, and the other uses them to get a loan. That said, biased results can sometimes marred suggestions of fairness and nondiscrimination. Thus, this blog examines the types of biases present in AI, which are data bias and algorithmic bias, before explaining why good governance is crucial to managing such disparities. Bias, as applied to AI, thus refers to a complicated technique of prejudice enshrined in decisions made by these systems and wherein individuals are treated prejudicially based on factors including but not limited to gender, race, and other aspects of socio-economic differentiation.</p><p><strong>Understanding Data Bias</strong></p><p>Bias in data happens when AI datasets contain prejudice or poorly sample or undersample some demographics. The second challenge mentioned in the work of Barocas et al. (2019) shows how datasets and data-gathering mechanisms could be biased, including non-random sampling of data, historical bias data, or actual lack of minority representation, resulting in AI methods that merely reinforce social inequalities. For instance, imagine a facial recognition system trained mainly from lighter-skinned people. That is why, for example, it may have lower accuracy when recognizing people from amongst those with dark skin tones, thus providing users with wrong output data and can lead to serious social consequences (Buolamwini &amp; Gebru, 2018). Preprocessing and data aggregation techniques reduce data bias by introducing inclusion and diversity policies as the training dataset’s collection and curation mechanisms.</p><p><strong>Understanding Algorithmic Bias</strong></p><p>In contrast, algorithmic bias occurs due to the inherent features of AI algorithms underpinning them, as well as resulting from the underlying interactions of the algorithms. Particular design elements may be biased toward particular users based on characteristics seen during algorithm construction. In other words, the learning paradigms within the designed algorithms reflect optimization goals that are racially and socially biased. In contrast to decision feedback provided at the end and exogenous to the original process, the open nature of feedback in these systems means positive feedback can entrench distortions and make their remediation harder.</p><p><strong>The Need for AI Governance</strong></p><p>The nature of data and algorithmic bias is complex and thus requires organizations to develop AI governance frameworks that specifically cover issues on fairness in the development of new AI and ensure that there will be repercussions for biases that may be identified in the algorithm. By governance structures, one should mean such measures and standards for preventing bias from being introduced when starting the AI development process and during its implementation and usage. Identifying practical guidelines includes regularly using various data in training and auditing patterns and procedures to eliminate or at least determine existing prejudice (Jobin et al., 2019). It is also fitting to form multi-disciplinary governance teams that involve data scientists, ethicists, and community members to help organizations recognize gaps in bias and develop workarounds concerning AI technology.</p><p><strong>Key Components of AI Governance for Bias Mitigation</strong></p><p>Eliminating bias in AI systems ensures that the governance frameworks comprise several essential elements for formulating fast, actionable policies, ethical frameworks, and accountability. By adopting these components, the following benefits are likely to be realized apart from addressing bias: firstly, responsibility and accountability are created for the growth of AI.</p><p><strong>1.</strong> <strong>Policies and Regulations</strong></p><p>AI governance starts with good, solid policies and regulations implemented correctly — as some of the most fundamental requirements. Such policies should focus on the data-gathering phase of implementing AI, the Algorithm designing stage, and the Implementation phase of the AI technologies. Based on the empirical work by Jobin et al. (2019), ethical assessments should, in turn, become obligatory and should encompass documentation procedures, including the pertinent bias risks. Some of the guidelines could be, for example, the proposal that before applying AI models, the bias impact assessments must be conducted, and the decision-making processes used should be audited for bias. Regulatory bodies should also impose methods of ensuring compliance with the regulation to make the organizing bodies accountable for AI technology.</p><p><strong>2.</strong> <strong>Ethical Frameworks</strong></p><p>Ethical frameworks provide directions for the proper application of Artificial Intelligence. They put measures such as fairness, accountability, and transparency — which are fundamentals of the ethical AI practice (O’Neil, 2016). O’Neil (2016) shows that ethical considerations should concern a variety of issues, and one of them is the problem of discrimination and social effects. For instance, frameworks may be chosen where the focus is the protection of the user’s privacy, but at the same time, users with disabilities will have equal access to these AI systems. It can also give developers direction regarding ethical issues and give them input as they develop, build, test, and implement AI solutions.</p><p><strong>3.</strong> <strong>Accountability Mechanisms</strong></p><p>Performance accountability systems are essential for demanding that organizations take responsibility for the AI systems they create. This may embrace hiring committees to constantly supervise the system and routine audits to check AI outcome bias. Impact assessments enable stakeholders to learn the consequences of the artificial intelligence system on different groups of people. Thus, organizations can give the stakeholders more knowledge on the issues connected with bias in AI technologies and report more frequently.</p><p>Also, accountability assumes repercussions, which must be placed on developers and other organizations with reckless disregard for ethical standards when developing AI systems. Therefore, A few checks could be deployed to prevent a lack of attention to bias since it reestablishes the steps toward responsible AI. By adopting all these accountability mechanisms, stakeholders in a given organization can cultivate a regime within which they would respond to any possibility of bias.</p><p><strong>Implementing Effective Governance Strategies</strong></p><p>Finally, numerous organizations need proper governance policies to incorporate these elements into artificial intelligence systems. Thus, complicated regulations also allow for much more effective work on creating better practices to govern when it is necessary to attract as many stakeholders as possible. This way, the correct governance of AI as a framework encompassing the datasets and algorithms used contributes to understanding and handling the ethical incidences bound to talent in AI technology and passing on healthy results to every technology user.</p><p><strong>1.</strong> <strong>Fostering Collaboration and Continuous Improvement</strong></p><p>Evaluating AI governance involves elements such as encouraging partnerships between individuals such as researchers, developers, ethicists, policymakers, and AI affectation populations. That way, various viewpoints can be nurtured, and AI systems can be applied to address the issue appropriately. It is possible to learn from the primary sources of the communities targeted by bias and come up with ways to counter this problem. When end-users are involved in the creation process, known as participatory design, implicit biases may be uncovered, and the end-user’s experience might be improved.</p><p><strong>2.</strong> <strong>Education and Training</strong></p><p>To supplement the preceding, training and Educational work among practitioners in the AI sector is the way forward to improving the existing governance in Artificial intelligence. The International Society for Ethical AI (ISE) 2021 partners with the AI Multi-Disciplinary Education (AMDE) program to create broadly based and holistic ethical, biased, and inclusive AI education. These programs seek to provide a person with working skills and the ability to respect social context and ethical considerations. Educational sessions, which include workshops, seminars, and knowledge-sharing sessions, are emphasized to ensure these principles focus on the development of AI while respecting responsibility and inclusiveness.</p><p><strong>3.</strong> <strong>Monitoring and Evaluation</strong></p><p>Specific attention to reinforcement learning and repeated testing of the models is valuable to address bias issues emerging in further operation. The policies should identify key performance indicators to measure bias within and set routines for outcome evaluations against differentiation in user characteristics. The concepts of ‘fair-by-design’ and ‘audit-by-design’ can help the organization introduce constant bias checks during its AI life cycle phases (Kim et al., 2019). Such constant monitoring assists in the early detection of bias and ensures organizations and the field continuously improve their models for enhancing user usability.</p><p><strong>The Role of Technology</strong></p><p>Apart from governance and organizational measures, some solutions stem from applying technology. The use of tools and methods like, for example, fairness-aware algorithms and bias detection software can act as preventive measures to overcome potential problems within an organization’s systems. These advantages mean the technologies can reveal emerging data or results biases and help intervene quickly.</p><p><strong>Conclusion: A Call to Action</strong></p><p>Bias is a never-ending process that requires a systemic solution that includes well-regulated solutions, teamwork, awareness, and many solutions based on sound technology. In this way, ensuring the proper understanding of fairness means achieving entirely justified AI utilization that will create the values of the entire population. The problem of bias, therefore, becomes pertinent as rapid frames in the AI development cycle are made minimally bias-free. There is a need for practice and policy synergy across the technology spectrum to guarantee that AI is a tool that empowers, enfranchises, and is just, not one that amplifies bigotry. Including diverse groups of people is a significant problem when using artificial intelligence in different sectors of society. However, if communities and states are interested in enhancing the systems, it is possible to work together to develop proper AI tools for the future.</p><p><strong>References</strong></p><p>Barocas, S., Hardt, M., &amp; Narayanan, A. (2019). <em>Fairness and Machine Learning: Limitations and Opportunities</em>. Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency. <a href="https://fairmlbook.org/">https://fairmlbook.org/</a></p><p>Buolamwini, J., &amp; Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. <em>Proceedings of the 2018 Conference on Fairness, Accountability, and Transparency,</em> 77–84. <a href="https://doi.org/10.1145/3287560.3287598">https://doi.org/10.1145/3287560.3287598</a></p><p>Jobin, A., Ienca, M., &amp; Andorno, R. (2019). The global landscape of AI ethics guidelines. <em>Nature Machine Intelligence</em>, 1(9), 389–399. <a href="https://doi.org/10.1038/s42256-019-0088-2">https://doi.org/10.1038/s42256-019-0088-2</a></p><p>Kim, P. T., et al. (2019). Fairness: A user-centric approach. In <em>Proceedings of the Conference on Fairness, Accountability, and Transparency</em> (pp. 137–145). <a href="https://doi.org/10.1145/3287560.3287594">https://doi.org/10.1145/3287560.3287594</a></p><p>O’Neil, C. (2016). <em>Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy</em>. Crown Publishing Group. <a href="https://dl.acm.org/doi/10.5555/3002861">https://dl.acm.org/doi/10.5555/3002861</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b2894657078c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/mitigating-bias-in-ai-systems-b2894657078c">Mitigating Bias in AI Systems</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Designing GEN AI Application Integrations]]></title>
            <link>https://medium.com/design-bootcamp/designing-gen-ai-application-integrations-f41d133c89e3?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/f41d133c89e3</guid>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Tue, 25 Feb 2025 23:57:36 GMT</pubDate>
            <atom:updated>2025-02-26T22:14:59.593Z</atom:updated>
            <content:encoded><![CDATA[<p>This blog aims to discuss integration patterns for GEN AI applications, emphasizing how LLM models can be best integrated into an application to provide optimal results.</p><p><strong>Overview</strong></p><p>The continuous implementation of generative AI (GEN AI) in numerous applications has become a revolutionary element in several fields. LLMs like GPT4 and Google Gemini have proven to comprehend and generate human language text accurately; hence, they are essential for improving work/output, decision-making, and the user interface (Bustos et al., 2024; Malhotra, 2024).</p><p>On the other hand, generative AI is an AI system that can craft new material, such as text, images, or music. Applying these models in IT applications entails solution ideas, fine-tuning state SET models, ensuring massive implementation of these models, and responsibility aspects (MDPI, 2024; NTT DATA, 2024). In this blog, we analyze integration patterns for GEN AI applications, opportunities for their realization, and critical risks.</p><p><em>a. Key Concepts and Techniques</em></p><p>Before understanding how it fits into GEN AI, familiarizing with some concepts and techniques will be helpful. These are pre-training, fine-tuning, prompt engineering, and retrieval augmented generation (RAG). The pre-training phase involves understanding the extensive data on which the model is trained. You get fine-tuning by training it on a smaller dataset of only the task in question or a particular domain (IEEE Xplore, 2024). An example is the engineering of efficient prompts that would lead to the intended output in the model, known as prompt engineering. Then, the retrieval augmented generation (RAG) integrates retrieval with generation to improve the model (Forbes, 2024).</p><p><em>b. Framework for Integration</em></p><p>To integrate GEN AI into IT applications, we need to build a structured method to contain entry points, pre-processing the input data, inference, modification of the output data, and presenting the end products (Bustos et al., 2024). Interfaces, however, describe how the model can be accessed or used in an application, such as API or software-loaded models (Leanpub, 2024). Impulsive pre-processing is done immediately before feeding the data into the model. Inference produces the result by using input data comprehensively (JK Tech, 2024). At the same time, post-processing involves adding other features to the generated output to meet certain specific qualities. The final aspect is called presentation or, in other words, how the output is utilized or displayed within the application (Sciendo, 2024).</p><figure><img alt="Illustration of LLM Framework" src="https://cdn-images-1.medium.com/max/1024/0*6lylbrsfg2jg5nBX.png" /></figure><p><em>c. Practical Applications</em></p><p>The utilization of GEN AI entails numerous and unlimited domains, including summarization, metadata extraction, intent classification, and question answering (MDPI, 2024; NTT DATA, 2024). For instance, Gen AI can intelligently identify the preferable quantities to produce in such plants by enhancing production resource allocation without escalating costs considerably in the manufacturing industry. In transportation logistics, since GEN AI can analyze the best routes and schedules in which delivery vehicles may operate, transport costs can be cut down (Forbes, 2024). In terms of marketing, GEN AI is capable of supporting how the marketing budget can be best divided in terms of marketing ROI.</p><p><em>d. Ethical Considerations</em></p><p>On one hand, the integration of GEN AI provides various opportunities; on the other hand, ethical concerns such as bias reduction, data privacy, and surveillance do exists. Bias management focuses on making the model’s performance accurate and free from prejudices; however, data confidentiality safeguards any sensitive data used for designing and running the model. It is required to follow changes in the model’s performance over time and recognize any problems that might occur.</p><p><strong>Integration Patterns for GEN AI</strong></p><p>Several integration patterns can be used to integrate large language models into IT applications (Bustos et al., 2024). These patterns provide a framework to facilitate extensibility, good interactions, high performance, and reusable design. Some of the popular integration patterns include:</p><p><em>a. API-based Integration</em></p><p>This is the most straightforward and frequently used technique for integrating GEN AI into applications through API. This pattern relies upon the application programming interface provided by AI service providers to determine the working capability of large language models. As a service, developers can pass requests to the API specifying the input they want for a specific generation task, and the API returns a response containing the generated output (IEEE Xplore, 2024). Thus, this pattern is highly flexible and can be brought into operation quickly without incurring the need to carry out large-scale infrastructural changes.</p><p><em>b. Native Integration</em></p><p>The second form of integration can be called putting the language models right inside the application. Nonetheless, this approach works by downloading the model and deploying it within the app’s infrastructure. Compared to the previous method, this one can be characterized by higher initial costs of implementation and even the necessity of additional infrastructure; however, it provides several advantages, ranging from lower latency to having much greater control over the work of the model (Leanpub, 2024).</p><p><em>c. Hybrid Integration</em></p><p>The semi-integrated model integrates API-based and embedded model integration elements. In this pattern, the application uses APIs for some operations, and others, it employs embedded models (Sciendo, 2024). It is flexible rules for using GEN AI in organizations based on requirements and restrictions to achieve optimal organizational performance.</p><p><em>d. Cloud-based Integration</em></p><p>Cloud integration is a process where GEN AI models are hosted in cloud environments and accessed through cloud services (JK Tech, 2024). This pattern is because it is highly flexible, meaning organizations can change the extent of their use depending on the situation. It also limits the demands on an organization’s on-premises equipment and facilities, benefiting organizations lacking substantial IT infrastructure.</p><p><strong>Case Studies</strong></p><p><em>Case Study 1: Improving Customer Service Using GEN AI</em></p><p>An international online shopping firm utilized GEN AI in its customer relation service to enhance the quality of response and, thus, customer satisfaction. In the form of an API integration pattern, the company employed a large language model capable of answering frequently asked questions by consumers (Upwork, 2024). This integration helped reduce the workload on human agents and offered better performance in handling complex issues as a general gain.</p><p><em>Case Study 2: Optimizing Content Creation</em></p><p>The content marketing firm used a mixed integration approach to support its content development. The first articles and drafts of social media posts were created with the help of integrating a gigantic language model into the firm’s content management system. At the same time, it employed cloud services for more specific works like keyword optimization and sentiment analysis (Forbes, 2024). They adopted this approach to streamline the creation of content, which was a milestone in helping the firm generate more significant volumes of content in a shorter time.</p><p><em>Case Study 3: Improving Medical Diagnostics</em></p><p>A healthcare organization includes GEN AI technology in all its diagnostic equipment to help doctors diagnose medical conditions (Lark, 2024). API-based integration pattern established the organization’s connectivity with a vast language model trained on medical publications to offer diagnosis recommendations relative to the patient’s symptoms and history. These two integrated solutions enable quicker and more accurate diagnoses, improving the patient’s health status.</p><p><strong>Future Directions</strong></p><p>GEN AI integration into applications has remained an area of improvement as new models are released and new methods of incorporating models are developed. Possible future works include developing finer-grained models suitable for particular industry segments, further enhancing strategies for eliminating the risk of biased model deployment, and defining integration patterns that would provide increased flexibility and scalability (MDPI, 2024).</p><p><strong>Conclusion</strong></p><p>Integrating large language models into applications brings about many opportunities to improve efficiency, make better decisions, and improve the user experience. Exploring several integration patterns, one can learn how to introduce Gen AI into an organization and bring vital improvements. However, the ethical factors and the latest trends should be reviewed and considered so that these effective technologies can be utilized meaningfully and ethically.</p><p><strong>References</strong></p><p>Analytics Insight. (2024). <em>Ethics of generative AI: Implications and considerations</em>. Retrieved from <a href="https://www.analyticsinsight.net/generative-ai/ethics-of-generative-ai-implications-and-considerations">https://www.analyticsinsight.net/generative-ai/ethics-of-generative-ai-implications-and-considerations</a></p><p>Arlington Public Library. (2024). <em>Generative AI application integration patterns integrate…</em>. Retrieved from <a href="https://libcat.arlingtonva.us/ExternalEContent/361273">https://libcat.arlingtonva.us/ExternalEContent/361273</a></p><p>Bustos, J. P., Soria, L. L., &amp; Arsanjani, A. (2024). <em>Generative AI application integration patterns: Integrate large language models into your applications</em>. Packt Publishing.</p><p>Forbes. (2024). <em>Ethical considerations for generative AI</em>. Retrieved from <a href="https://www.forbes.com/councils/forbestechcouncil/2024/02/14/ethical-considerations-for-generative-ai/">https://www.forbes.com/councils/forbestechcouncil/2024/02/14/ethical-considerations-for-generative-ai/</a></p><p>IEEE Xplore. (2024). <em>Generative AI application integration patterns: Integrate large language models into your applications</em>. Retrieved from <a href="https://ieeexplore.ieee.org/document/10769230">https://ieeexplore.ieee.org/document/10769230</a></p><p>JK Tech. (2024). <em>Ethical considerations and challenges of using generative AI</em>. Retrieved from <a href="https://jktech.com/blogs/ethical-considerations-and-challenges-of-using-generative-ai/">https://jktech.com/blogs/ethical-considerations-and-challenges-of-using-generative-ai/</a></p><p>Lark. (2024). <em>Ethical issues in generative AI</em>. Retrieved from <a href="https://www.larksuite.com/en_us/topics/ai-glossary/ethical-issues-in-generative-ai">https://www.larksuite.com/en_us/topics/ai-glossary/ethical-issues-in-generative-ai</a></p><p>Leanpub. (2024). <em>Generative AI application integration patterns</em>. Retrieved from <a href="https://leanpub.com/generativeaiapplicationintegrationpatterns">https://leanpub.com/generativeaiapplicationintegrationpatterns</a></p><p>Malhotra, Y. (2024). <em>Generative AI application integration patterns: Integrate large language models into your applications</em>. Packt Publishing.</p><p>MDPI. (2024). Ethical challenges and solutions of generative AI: An interdisciplinary perspective. <em>Informatics, 11</em>(3), 58. <a href="https://doi.org/10.3390/informatics11030058">https://doi.org/10.3390/informatics11030058</a></p><p>NTT DATA. (2024). <em>Perspective: Ethical considerations of generative AI</em>. Retrieved from <a href="https://www.nttdata.com/global/en/-/media/nttdataglobal/1_files/insights/reports/generative-ai/ethical-considerations-of-genai/ethical-considerations-of-generative-ai.pdf">https://www.nttdata.com/global/en/-/media/nttdataglobal/1_files/insights/reports/generative-ai/ethical-considerations-of-genai/ethical-considerations-of-generative-ai.pdf</a></p><p>Sciendo. (2024). <em>Generative AI application integration patterns</em>. Retrieved from <a href="https://sciendo.com/book/9781835887615">https://sciendo.com/book/9781835887615</a></p><p>Upwork. (2024). <em>6 ethical considerations of artificial intelligence</em>. Retrieved from <a href="https://www.upwork.com/resources/ai-ethical-considerations">https://www.upwork.com/resources/ai-ethical-considerations</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f41d133c89e3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/designing-gen-ai-application-integrations-f41d133c89e3">Designing GEN AI Application Integrations</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA["Security by Design in Identity Governance: Why IAM Architectures Must Be Secure by Default]]></title>
            <link>https://medium.com/iga-experts/designing-security-for-software-products-69e3dd43213f?source=rss-7999a8b5ed7d------2</link>
            <guid isPermaLink="false">https://medium.com/p/69e3dd43213f</guid>
            <category><![CDATA[design-thinking]]></category>
            <category><![CDATA[software-devlopment]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Anant Wairagade]]></dc:creator>
            <pubDate>Sun, 09 Feb 2025 07:06:38 GMT</pubDate>
            <atom:updated>2026-05-08T16:58:08.852Z</atom:updated>
            <content:encoded><![CDATA[<h3>Security by Design in Identity Governance: Why IAM Architectures Must Be Secure by Default</h3><p>An approach to building security into software products without explicitly thinking about it.</p><p>Security being an afterthought in development is an all-too-common tendency. This is why there is a huge emphasis on building security into software products. But how you go about it matters.</p><p>Explicitly thinking about security as you build your code could easily cause it to fall down the pecking order, After all, something has to work before you can protect it, right? Well, not necessarily.</p><p>This is where security by design comes in — an approach to building security into software products without explicitly thinking about it. In this article, we’ll cover what this approach is, how it works, and why it is the best way to secure your software.</p><h3>What does security by design mean?</h3><p>This is an approach where security is integrated throughout the software development lifecycle, starting from the design phase. In this proactive approach, security is considered from the outset, not added later as a patch. In other words, security is built-in, allowing you to rely on your code rather than putting all your faith in external resources to keep your software secure.</p><p>The goal is to build software that is inherently secure, minimizing vulnerabilities and the likelihood of a successful attack.</p><p>Before we see how to go about how to achieve security by design, let’s first go over how you can start thinking of how to implement it.</p><h3>Viewing security as a concern rather than a feature</h3><p>By default, software development is feature-driven. Developers place their primary focus on what the software can do. This is supported by software development methodologies like the Rational Unified Process (RUP) which focuses on the use cases of a piece of software.</p><p>Consequently, it’s no surprise that developers tend to view security as a set of features to be implemented. This shouldn’t be the case, as describing security this way often misses the point.</p><h3>A practical example of viewing security as a set of features</h3><p>Let’s take a look at an example:</p><p>Suppose you want to build a photo gallery website where users can upload their photos and have them stay confidential. If you view security as a set of features, you would have a user story similar to this “As a user I want to log into the website and access my photos”.</p><p>If you implement a login page, you’ve met the objectives of this user story. However, you’ve not addressed security concerns because if someone else stumbles on the download link, they can still download the user’s photos.</p><figure><img alt="illustration of  unauthorized access by a user" src="https://cdn-images-1.medium.com/max/690/0*30j0LUkip2bKpYJP" /></figure><p>A better way to do this would be to view security as a concern, and the concern here is that there should be no way for a user to access photos without logging in. This will direct you to look into the different ways to mitigate this concern beyond a simple username and password.</p><p>In essence, to get real security, you need to get away from thinking about security as a set of features rather than a concern. Then finding ways to mitigate these concerns, limiting vulnerabilities in the long run.</p><p>One way to do this is what we refer to as the traditional approach to software security.</p><h3>The traditional approach to building security into software and its shortcomings</h3><p>In this approach, you view security as a top priority when writing and developing your code to mitigate security vulnerabilities. Everyone involved in development and testing should be experienced or trained in cyber security.</p><p>They should know about things such as:</p><ul><li>Cross-site scripting (XSS).</li><li>Vulnerabilities in low-level protocols.</li><li>The <a href="https://owasp.org/www-project-top-ten/">OWASP top 10</a>, and so on.</li></ul><p>Knowing this allows them to make decisions about software security with every block of code they write.</p><p>To better understand this approach, here is what it might look like:</p><p>Suppose you have an object where a username is displayed on some page. As expected, if the user name value is assigned directly without validation, it opens up an entry point for an XSS attack.</p><figure><img alt="code snippet illustrating XSS cyber attack" src="https://cdn-images-1.medium.com/max/991/0*XXtahGgfBdaSpL2b" /></figure><p>To mitigate this with the traditional approach, you could introduce security-focused input validation. For example, by adding code that checks that parameters aren’t null and pulling a security library to validate user input for possible XSS attacks.</p><h3>Limitations of the traditional method</h3><p>In all fairness, this way of handling security is simple and thus quite common. However, it is also problematic for several reasons, such as:</p><ul><li>Developers and testers need to explicitly think about security vulnerabilities while focusing on solving business functionality at the same time. This creates a conflict between focusing on security and functionality.</li><li>It assumes everyone involved is a security expert, which is impractical. The software developers create will reflect their level of security.</li><li>It also assumes that the developer can think of all possible vulnerabilities that could occur now and in the future, which is impossible.</li></ul><p>Because of these limitations, this approach struggles to create software secure enough for production. To achieve true security, we need to look elsewhere — and this is where security by design comes in.</p><h3>Driving security through design</h3><p>This involves focusing on software design instead of explicitly thinking about security as you write your code. By focusing on design, we mean that you’re always aiming for the highest possible standards with what you create. In doing so, you achieve a high degree of software security without needing to constantly and explicitly think about it.</p><p>Let’s go back to our example to understand how this works.</p><h3>Example: Creating a secure domain through design</h3><p>To approach this example by focusing on good design, you’ll first need to define the meaning of a username in the context of the current application. You’ll conclude that the username can only contain letters (both upper and lower case) and numbers (0–9).</p><p>You’ll also say that the username should be 4–40 characters long because this is how long a username typically is. Next, you can ask yourself whether &lt; or &gt; characters are usually part of usernames. Then decide to exclude them, not because they can be part of an XSS attack but because a username usually doesn’t have these characters.</p><p>To sum it up, a user’s input has to pass through 3 checks:</p><figure><img alt="illustration of eliminating potential XSS security vulnerability by following secure design practice for domain check" src="https://cdn-images-1.medium.com/max/650/0*1HSTaSNgm6-5WZTN" /></figure><p>By focusing on good design, you can start eliminating security vulnerabilities even before you start thinking about security. Since “&lt;script&gt;alert(‘XSS’)&lt;/script&gt;” is no longer a viable input, you’ve protected your domain against the XSS vulnerability we discussed earlier.</p><p>What’s more, with a design-first approach, you learn more about your domain and create a more precise domain model.</p><h3>Code constructs promoting security</h3><p>There are several design-focused strategies you can use to solve security problems you might encounter in your daily work. They include immutability, fail-fast contracts, and validation which we’ve already covered above. So now let’s go over immutability and fail-fast contracts.</p><h4>Immutability</h4><p>Immutability simply means that the state of your object shouldn’t change. While this seems trivial, it makes all the difference when it comes to security. Immutable objects can be safely shared between the sub-processes of a program and make data highly available. This is important in preventing denial-of-service attacks.</p><h4>Fail-fast contracts</h4><p>Many security problems arise because one part of a system assumes that another part takes responsibility for something when the other part also assumes the opposite. Fortunately, you can think about design in terms of preconditions and contracts, which enables you to specify which part of a design takes on which responsibility.</p><p>Design contracts specify the preconditions required for an object to work as intended and the postconditions for how the object changes after a method is completed. In a fail-fast contract design, a system will halt its execution immediately and raise an error as soon as it detects a condition that violates its expectations (e.g., invalid input). This is important in safeguarding against unexpected situations.</p><p>In essence, a strong focus on design allows you to create code that is more secure compared to the traditional approach. Now, let’s dive deeper into the advantages of this method and understand why it is better than the traditional approach.</p><h3>The advantages of driving security through design</h3><p>When you focus on design, security can become part of the development process naturally instead of being viewed as a forced requirement. As such, a design-first approach helps avoid many of the shortcomings of the traditional approach and has its own set of advantages. Here are some of the main ones.</p><h4>Design comes naturally to software developers</h4><p>Understanding the intricacies of software vulnerabilities is something most developers struggle with. But they typically understand and appreciate good design practices.</p><p>In fact, they take pride in creating good designs that serve their purpose well. If you can use design to achieve security, then security also comes naturally to your development processes.</p><h4>No conflict between functionality and security</h4><p>A huge downside with the traditional approach is that it decouples security with functionality, forcing the two to compete against each other. However, with security by design, you focus on creating high-quality designs and achieve security as you do. This means that functionality concerns and security become of equal priority.</p><h4>Non-security experts can naturally write secure code</h4><p>You don’t have to be a security expert to write secure code with this approach. How you design your software implicitly avoids insecure constructs, even before you explicitly start thinking about security.</p><h3>Don’t forget about security</h3><p>Throughout this article, we’ve emphasized how to not think about security but still get it anyway. But we shouldn’t overlook it entirely. You’ll need to combine the traditional and design-first approaches to get the best results.</p><h4>Conduct code security reviews</h4><p>If you aren’t conducting code reviews already, we highly recommend you do. Code reviews are a great way to find possible design flaws that could be security vulnerabilities.</p><p>These are more like code reviews except they’re aimed at reviewing the security properties of the code. A great way to do this would be to have a checklist with questions such as:</p><ul><li>Are security-related HTTP headers used correctly?</li><li>How often are passwords rotated in the system?</li><li>How are passwords protected and stored in the system?</li></ul><p>This is just a general example. You can create a checklist that suits your applications.</p><h4>Run security penetration tests</h4><p>Security vulnerabilities can exist anywhere, for instance, in your application code or the deployment pipeline. Security penetration tests, or simply pen tests, provide a way for you to challenge your design and see if it is truly impenetrable.</p><p>Your software will evolve, and you’ll add new business functions to it. These new functions could cause exploitable vulnerabilities. This makes a great case for running pen tests regularly, so you can detect and solve them before they become a bigger problem down the line.</p><p>Apart from these, there is so much more you can do here, including studying security trends to stay on top of them and developing a security incident mechanism. What you can take away from this is that the design-first approach doesn’t render some traditional methods obsolete.</p><h3>Conclusion</h3><p>Security by design is an approach for improving your software’s security without explicitly focusing on it. You work on good design practices that naturally incorporate secure code design practices. It is a way to consolidate functionality and security so that neither one of them trumps the other. While this approach encourages focusing less on security while designing software, you shouldn’t abandon security measures altogether. You can combine these methods for the best results.</p><h3>References</h3><ol><li><strong>Howard, M., &amp; Lipner, S.</strong> (2006). <em>The Security Development Lifecycle</em>. Microsoft Press.</li><li><strong>Anderson, R.</strong> (2020). <em>Security Engineering: A Guide to Building Dependable Distributed Systems</em> (3rd ed.). Wiley.</li><li><strong>National Institute of Standards and Technology (NIST).</strong> (2018). <em>Systems Security Engineering: Considerations for a Multidisciplinary Approach in the Engineering of Trustworthy Secure Systems</em> (NIST SP 800–160, Volume 1). <a href="https://doi.org/10.6028/NIST.SP.800-160v1">https://doi.org/10.6028/NIST.SP.800-160v1</a></li><li><strong>OWASP.</strong> (2021). <em>OWASP Secure Design Principles</em>. <a href="https://cheatsheetseries.owasp.org/cheatsheets/Secure_Design_Principles_Cheat_Sheet.html">https://cheatsheetseries.owasp.org/cheatsheets/Secure_Design_Principles_Cheat_Sheet.html</a></li><li><strong>Cybersecurity and Infrastructure Security Agency (CISA).</strong> (2023). <em>Secure by Design: Principles and Guidelines for Developers</em>. <a href="https://www.cisa.gov/securebydesign">https://www.cisa.gov/securebydesign</a></li><li><strong>Johnsson, D., Deogun, D., &amp; Sawano, D.</strong> (2019). Secure by Design. Manning Publications.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69e3dd43213f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/iga-experts/designing-security-for-software-products-69e3dd43213f">&quot;Security by Design in Identity Governance: Why IAM Architectures Must Be Secure by Default</a> was originally published in <a href="https://medium.com/iga-experts">IGA Experts</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>