<?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 Pooja Kamath on Medium]]></title>
        <description><![CDATA[Stories by Pooja Kamath on Medium]]></description>
        <link>https://medium.com/@poojapkamath?source=rss-4b73901802bb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*CCinYsEsNbxu0dgqHtRY6A@2x.jpeg</url>
            <title>Stories by Pooja Kamath on Medium</title>
            <link>https://medium.com/@poojapkamath?source=rss-4b73901802bb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:28:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@poojapkamath/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[Accessible Error Contracts: Designing the userMessage Pattern in RAML and OAS]]></title>
            <link>https://medium.com/@poojapkamath/accessible-error-contracts-designing-the-usermessage-pattern-in-raml-and-oas-f84b7fd44558?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/f84b7fd44558</guid>
            <category><![CDATA[testing]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[raml]]></category>
            <category><![CDATA[web-accessibility]]></category>
            <category><![CDATA[mulesoft]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 18:52:40 GMT</pubDate>
            <atom:updated>2026-08-11T18:52:40.035Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Accessibility Series: 2</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W5aHiuHl_WvlRUwa" /><figcaption>Photo by <a href="https://unsplash.com/@brett_jordan?utm_source=medium&amp;utm_medium=referral">Brett Jordan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In the <a href="https://medium.com/@poojapkamath/designing-accessible-integrations-855d51550102">last post</a>, we made the case that accessibility is mostly an Experience-layer concern, and that one of the cheapest, highest-leverage fixes is a userMessage field on your error responses — a string meant to be read aloud or displayed inline, distinct from your internal error code.</p><p>That’s easy to say in a paragraph. It’s a different thing to actually specify, transform, test, and version. This post is the implementation guide: how to model the field in RAML or OAS, how to populate it in DataWeave without duplicating logic across every flow, how to verify it actually works with a screen reader, and how to change it later without breaking every consumer who built against the old contract.</p><h3>Start with the schema, not the flow</h3><p>The temptation is to add userMessage ad hoc, inside whichever flow happens to throw the error. Resist it. If userMessage isn&#39;t part of your published error schema, it&#39;s not a contract — it&#39;s a convention that will drift the first time a new developer touches the project.</p><p><strong>RAML</strong> — define a reusable error type once, in a library, and reference it everywhere:</p><pre>#%RAML 1.0 Library<br>types:<br>  ApiError:<br>    type: object<br>    properties:<br>      status:<br>        type: integer<br>        example: 400<br>      error:<br>        type: string<br>        description: Machine-readable error code for logging and client branching<br>        example: VALIDATION_ERROR<br>      message:<br>        type: string<br>        description: &gt;<br>          Human-readable string, safe to read aloud or display inline.<br>          Must reference the specific field and the corrective action.<br>        example: &quot;The policy number field is required and was left blank.&quot;<br>      field:<br>        type: string<br>        required: false<br>        example: policyNumber<br>      userAction:<br>        type: string<br>        required: false<br>        description: One clear instruction the user can act on immediately.<br>        example: &quot;Enter your policy number in the format XXX-XXXXXXX.&quot;</pre><p><strong>OAS 3.x</strong> — the equivalent as a reusable component:</p><pre>components:<br>  schemas:<br>    ApiError:<br>      type: object<br>      required: [status, error, message]<br>      properties:<br>        status:<br>          type: integer<br>          example: 400<br>        error:<br>          type: string<br>          description: Machine-readable error code<br>          example: VALIDATION_ERROR<br>        message:<br>          type: string<br>          description: Human-readable, screen-reader-safe explanation<br>          example: &quot;The policy number field is required and was left blank.&quot;<br>        field:<br>          type: string<br>          example: policyNumber<br>        userAction:<br>          type: string<br>          example: &quot;Enter your policy number in the format XXX-XXXXXXX.&quot;</pre><p>Two things matter more than the YAML itself:</p><ul><li><strong>message and </strong><strong>error are not the same field wearing two hats.</strong> error is for log aggregation and client-side branching (if error === &#39;VALIDATION_ERROR&#39;). message is for a human. Conflating them is how you end up with &quot;error&quot;: &quot;ERR_4471_VALIDATION_FAILURE_POLICY_SVC&quot; being read aloud to a policyholder.</li><li><strong>Document the <em>intent</em> of the field in the schema description</strong>, not just its type. “Human-readable string, safe to read aloud” tells the next developer what not to break. A bare type: string doesn&#39;t.</li></ul><h3>Populating it without duplicating logic</h3><p>The failure mode here isn’t a missing field — it’s fifty flows each hardcoding their own version of a similar message, drifting in tone and completeness. Centralize message generation the same way you’d centralize any other cross-cutting concern: a DataWeave module.</p><pre>%dw 2.0<br>// modules/ErrorMessages.dwl<br>fun buildUserMessage(field: String, ruleType: String, format: String | Null = null) =<br>  ruleType match {<br>    case &quot;REQUIRED&quot; -&gt;<br>      { message: &quot;The $(field) field is required and was left blank.&quot;,<br>        userAction: &quot;Enter a value for $(field).&quot; }<br>    case &quot;FORMAT&quot; -&gt;<br>      { message: &quot;The $(field) field isn&#39;t in the expected format.&quot;,<br>        userAction: &quot;Enter $(field) in the format $(format default &#39;&#39;).&quot; }<br>    case &quot;RANGE&quot; -&gt;<br>      { message: &quot;The $(field) value is outside the allowed range.&quot;,<br>        userAction: &quot;Check the allowed range for $(field) and try again.&quot; }<br>    else -&gt;<br>      { message: &quot;There was a problem with the $(field) field.&quot;,<br>        userAction: &quot;Review $(field) and try again.&quot; }<br>  }</pre><p>Then a single error-handling scope, applied consistently across your Experience API’s flows, calls this module rather than each flow writing its own strings:</p><pre>%dw 2.0<br>import buildUserMessage from modules::ErrorMessages<br>output application/json<br>var validationResult = buildUserMessage(<br>  field: vars.failedField,<br>  ruleType: vars.failureType,<br>  format: vars.expectedFormat default null<br>)<br>---<br>{<br>  status: 400,<br>  error: &quot;VALIDATION_ERROR&quot;,<br>  message: validationResult.message,<br>  field: vars.failedField,<br>  userAction: validationResult.userAction<br>}</pre><p>This buys you two things beyond consistency: a single place to run copy through localization later, and a single place to fix tone or clarity issues once QA (or a real screen-reader test) flags them — instead of hunting through every flow that throws a 400.</p><h3>Testing it with an actual screen reader</h3><p>A userMessage field that reads fine in a JSON viewer can still fail in the browser. The gap is almost always in how the <em>frontend</em> wires the field to the DOM, not in the API itself — but if your Experience API team owns both, or reviews frontend integration, it&#39;s worth testing end to end at least once per pattern.</p><p>A lightweight process that doesn’t require a dedicated accessibility engineer:</p><ol><li><strong>Trigger the error in a real form</strong>, not Postman. Use NVDA (Windows, free) or VoiceOver (macOS, built in — Cmd+F5) and navigate to the field using only the keyboard.</li><li><strong>Check announcement, not just presence.</strong> The message needs to be wired to an aria-live region or associated with the field via aria-describedby — a userMessage sitting in the DOM but not referenced by either won&#39;t be announced at all. This is a frontend implementation detail, but it&#39;s the reason the API-level work exists, so it&#39;s worth verifying rather than assuming.</li><li><strong>Listen for redundancy and dead air.</strong> A common failure is the screen reader announcing the field label, then the generic browser validation message, then your custom userMessage — three overlapping announcements for one error. If you&#39;re rolling your own validation UI, suppress the browser default and let your userMessage be the only thing announced.</li><li><strong>Confirm the </strong><strong>userAction is actionable in isolation.</strong> Read only the userAction string out loud, with no other context. If it doesn&#39;t tell you what to physically do next, it&#39;s still written for a sighted user glancing at a red-bordered field.</li></ol><p>None of this requires a lab. Ten minutes with VoiceOver on your three or four highest-traffic error states catches most of what matters, and it’s a good addition to the same design review where you’re already asking the three accessibility questions from the last post.</p><h3>Versioning without breaking existing consumers</h3><p>userMessage fields evolve — you&#39;ll rewrite copy, add userAction to error types that didn&#39;t have it, or split a vague message into a more specific one. The risk is the same as any contract change: a client parsing message with a regex, or a test asserting exact string equality, breaks silently.</p><p>A few rules that keep this additive rather than breaking:</p><ul><li><strong>Adding </strong><strong>userAction to an error type that previously lacked it is safe.</strong> It&#39;s a new optional field; existing consumers ignoring it are unaffected. Adding <em>new optional fields</em> to ApiError never requires a version bump.</li><li><strong>Changing the wording of an existing </strong><strong>message is not safe if any consumer does exact-match testing on it — and some will</strong>, especially test suites that were written by copying the error string out of a bug report. Treat message <em>wording</em> changes like you&#39;d treat a UI copy change: fine to ship, but call it out in release notes for API consumers, the same way you&#39;d flag a UI copy change to a design team.</li><li><strong>Changing the <em>meaning</em> or <em>structure</em> of </strong><strong>field or </strong><strong>error is a breaking change</strong>, full stop — those are the machine-readable parts of the contract, and client-side branching logic depends on their stability. Version the API (or at minimum the error schema) if you&#39;re changing what error: &quot;VALIDATION_ERROR&quot; means or splitting it into more granular codes.</li><li><strong>Deprecate error codes the same way you’d deprecate an endpoint.</strong> If VALIDATION_ERROR is being split into VALIDATION_ERROR_REQUIRED and VALIDATION_ERROR_FORMAT, keep emitting the old code alongside the new one for a deprecation window, and document the migration the way you&#39;d document any other RAML/OAS breaking change.</li></ul><p>The underlying principle: error and field are the API contract in the traditional sense — client code branches on them, so change them like you&#39;d change any other contract. message and userAction are closer to localized UI copy that happens to live in your API — you should still be deliberate about changing them, but the bar is &quot;communicate the change,&quot; not &quot;version the API.&quot;</p><h3>Where this leaves you</h3><p>None of this is exotic. It’s a reusable RAML/OAS type, one DataWeave module instead of fifty inline strings, ten minutes with VoiceOver, and the same versioning discipline you already apply to every other field in your contract. The only new idea is treating message as a first-class part of the API design — something with its own quality bar — rather than whatever string happened to be convenient to throw.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f84b7fd44558" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Designing Accessible Integrations]]></title>
            <link>https://medium.com/@poojapkamath/designing-accessible-integrations-855d51550102?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/855d51550102</guid>
            <category><![CDATA[api-development]]></category>
            <category><![CDATA[mulesoft]]></category>
            <category><![CDATA[user-experience-design]]></category>
            <category><![CDATA[accessibility]]></category>
            <category><![CDATA[design]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 15:53:13 GMT</pubDate>
            <atom:updated>2026-07-28T15:53:13.231Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What API-Led Connectivity Owes the End User</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zB_MTqzOdZUe_qFk" /><figcaption>Photo by <a href="https://unsplash.com/@themephotos?utm_source=medium&amp;utm_medium=referral">Theme Photos</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Accessibility conversations in integration circles usually stop at the tooling. Is Anypoint Studio screen-reader friendly? Is API Console WCAG-compliant? Those are fair questions, but they miss a bigger one: <strong>are the integrations we build accessible to the people who eventually depend on them?</strong></p><p>An Experience API that returns a technically correct 400 error with no human-readable detail. A MuleSoft for Flow approval form with unlabeled fields. A monitoring alert that only fires as a color change on a dashboard. None of these break functionally. All of them quietly exclude someone. And because accessibility rarely shows up in an integration’s acceptance criteria, it’s one of the easiest things to ship broken.</p><p>This isn’t a call to add an accessibility checklist to your Jira board and move on. It’s a case for treating accessibility as an architectural concern — one that fits naturally into patterns integration teams already use.</p><h3>Accessibility is an Experience-layer problem, mostly</h3><p>API-led connectivity already gives you a natural boundary for this work. System and Process APIs deal in data contracts; Experience APIs deal in *people*. That’s not a throwaway distinction — it means the accessibility burden should concentrate almost entirely at the Experience layer, and that’s good news. You don’t need to retrofit every backend integration; you need to be deliberate about the layer that actually touches a human.</p><p>A few implications of that:</p><p>● <strong>System and Process APIs should stay accessibility-agnostic.</strong> Their job is correctness and reuse, not presentation. Don’t let UI concerns leak backward into your canonical data models — that’s the same discipline that keeps those layers reusable in the first place.</p><p>● <strong>Every Experience API needs an owner who thinks about the consumer, not just the consuming app.</strong> If three different channels (web, mobile, voice assistant) consume the same Experience API, accessibility requirements may differ per channel, and that logic belongs in the Experience layer, not scattered across client apps.</p><p>● <strong>Reused Experience APIs compound the problem or the fix.</strong> An inaccessible error-handling pattern baked into a widely reused Experience API propagates to every app that calls it. Get it right once, and you fix it everywhere at once — the same leverage that makes API-led connectivity valuable in general.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uTeLjtYT_KRUtYai7KtpFw.png" /><figcaption><strong>Designing Accessible Integrations</strong></figcaption></figure><h3>Pattern 1: Error responses that mean something to assistive technology</h3><p>Most API error handling is designed for developers debugging Postman calls, not for the systems downstream that translate those errors into what a user actually hears or reads.</p><p>Consider a validation failure on a claims-submission Experience API:</p><p>{</p><p>“status”: 400,</p><p>“error”: “VALIDATION_ERROR”,</p><p>“message”: “Invalid input”</p><p>}</p><p>That’s enough for a developer. It’s useless for a screen reader announcing an error to a policyholder filling out a form, because there’s no field reference and no actionable instruction. Compare:</p><p>{</p><p>“status”: 400,</p><p>“error”: “VALIDATION_ERROR”,</p><p>“message”: “The policy number field is required and was left blank.”,</p><p>“field”: “policyNumber”,</p><p>“userAction”: “Enter your policy number in the format XXX-XXXXXXX.”</p><p>}</p><p>The second version isn’t more expensive to produce in DataWeave — it just requires deciding, at the Experience API design stage, that message is a string meant to be read aloud or displayed inline, not a debug log line. This is a small contract change with an outsized downstream effect: it’s the difference between a screen reader saying “error” and a screen reader saying what to fix.</p><p><strong>Practical step:</strong> when you write your RAML/OAS error schema, add a userMessage field distinct from your internal error code, and make it a documented, tested part of the contract — not an afterthought filled in by whichever frontend team consumes the API.</p><h3>Pattern 2: MuleSoft for Flow and no-code forms</h3><p>MuleSoft for Flow puts integration-building in the hands of business users, which is exactly where accessibility mistakes tend to slip in — not from malice, but because form-building tools optimize for speed, not semantics.</p><p>A few habits worth building into your Flow templates or governance guidelines:</p><p>● <strong>Label every field explicitly</strong>, even when a placeholder makes the label seem redundant visually. Placeholder text disappears on focus and isn’t reliably read by all assistive tech.</p><p>● <strong>Group related fields with clear section headings</strong> rather than relying on visual spacing alone — spacing conveys nothing to a screen reader.</p><p>● <strong>Don’t encode meaning in color alone.</strong> A required field marked only with a red asterisk fails for colorblind users and for anyone using a screen reader, which won’t announce color. Pair color with text (“Required”) or an ARIA attribute if the form renders to HTML.</p><p>● <strong>Sequence matters.</strong> If a Flow-built approval workflow presents fields out of tab order, keyboard-only users navigate a form that doesn’t match its visual layout.</p><p>If your organization publishes reusable Flow templates through Exchange, this is a governance opportunity: bake accessibility checks into the templates once, and every business user who reuses them inherits compliant forms by default — the same reuse leverage that makes API-led connectivity effective, applied to low-code.</p><h3>Pattern 3: Notifications and alerts beyond the dashboard</h3><p>Anypoint Monitoring and most alerting integrations default to visual signals — a red tile, a status change on a dashboard. That works fine until the person watching the dashboard uses a screen reader, or isn’t watching a screen at all.</p><p>If your integration architecture pushes operational alerts downstream (to Slack, email, PagerDuty, or a custom notification API), a few adjustments make those channels accessible by default rather than by luck:</p><p>● Text-first alert payloads, with severity stated in words (“CRITICAL: Payment gateway timeout”) rather than conveyed only through color-coded UI in the receiving tool.</p><p>● Avoid relying on emoji or icons as the sole severity indicator in Slack/Teams messages — screen readers announce emoji inconsistently, and some convey no meaning at all out of context.</p><p>● If you’re building a custom notification API, include a structured severity and summary field separate from any formatted/HTML body, so downstream consumers that need plain text (SMS gateways, voice alerts, screen readers) aren’t stuck parsing markup.</p><h3>A lightweight way to start</h3><p>You don’t need a formal accessibility audit to make progress. Three questions, applied at Experience API design review, catch most of what matters:</p><p>1. <strong>If this response were read aloud by a screen reader with no visual context, would it make sense?</strong></p><p>2. <strong>Does anything here rely on color, icon, or spacing alone to convey meaning?</strong></p><p>3. <strong>Could someone complete this flow using only a keyboard, in the order the fields are meant to be filled?</strong></p><p>None of these require new tooling. They require adding one more lens to the same design reviews your Experience APIs already go through — the same discipline you’d apply to versioning, security, or backward compatibility.</p><h3>The reuse argument, one more time</h3><p>MuleSoft’s whole pitch for API-led connectivity has always been: build it once, well, and everyone downstream benefits. Accessibility fits that argument better than almost anything else in the platform. A well-designed error contract, a properly labeled Flow template, a text-first notification schema — these aren’t one-off fixes. They’re patterns that, once established at the Experience layer, get inherited automatically by every app, channel, and business user who reuses them.</p><p>That’s a lower bar than “run a full accessibility audit on every project.” It’s just: treat the people at the end of the API call as part of the contract, not an afterthought to it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=855d51550102" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Security Perimeter Stops at the AI Layer. That’s a Problem.]]></title>
            <link>https://medium.com/@poojapkamath/security-perimeter-stops-at-the-ai-layer-thats-a-problem-3fe714610bf0?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/3fe714610bf0</guid>
            <category><![CDATA[operations-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 22:39:39 GMT</pubDate>
            <atom:updated>2026-06-24T22:39:39.873Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Your AI is operating as designed, but it still needs a leash.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0d26LcFhRpx6d482UYroaw.jpeg" /></figure><p>I walk my dog every morning. He’s a genuinely good boy — calm around strangers, listens well, wouldn’t hurt a fly. But the second he spots a squirrel, a leaf blowing sideways, or another dog three blocks away, something primal kicks in. He doesn’t make a bad decision. He just <em>stops being mine to direct</em>.</p><p>So I leash him. Not because I distrust him. Not because he’s dangerous. Because the environment is unpredictable, and trust alone is not a control mechanism.</p><p>Enterprise AI is exactly the same problem. And right now, most organizations are letting their models run off-leash — confident they’ll be fine because they passed the intake review.</p><blockquote><em>“Compliance said the data never leaves the region. The model just replicated your proprietary decision logic and exposed it through an edge-node cache with zero access-control logging.”</em></blockquote><p>This is not hypothetical. This is Tuesday.</p><p>The brutal truth of modern enterprise architecture is simple: <strong>Your security perimeter stops at the AI layer. And that is a massive problem.</strong></p><h4>Are You Really Holding The Keys?</h4><p>There is a collision happening right now inside every enterprise deploying real-time AI. On one side: your data compliance team, guardians of GDPR, data sovereignty mandates, and regional residency requirements. On the other: your engineering team, shipping fast, integrating AI at the edge, pushing inference to distributed nodes to hit latency SLAs.</p><p>These teams are not enemies. But they are operating with fundamentally incompatible mental models of where risk lives.</p><p>Compliance thinks in terms of data location. Where does the data sit? Is it encrypted at rest? Is it stored in an approved region? Check, check, check. Done.</p><p>Engineering thinks in terms of data flow. What does the model need to run? What can we cache for performance? How do we get sub-100ms responses at the edge?</p><p>Neither is wrong. But together, they leave a gap wide enough to drive a regulatory violation through. While compliance is auditing storage, the AI is already <em>using</em> the data — processing it, reasoning over it, synthesizing proprietary context — in ways that no encryption key protects.</p><blockquote><strong><em>The Reality Check:</em></strong><em> Data-at-rest encryption offers exactly zero protection against data-in-use vulnerabilities inside an LLM or neural network context window.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zzHiBB0bW627pmtxScvuGw.png" /></figure><h4>Are You In Control?</h4><p>Data-at-rest encryption is a solved problem. It is also, increasingly, a decoy. Enterprises have spent a decade hardening their vaults. The data is locked. The perimeter is watched. The certificates are current. And the AI agent is still doing exactly what it wants with everything inside.</p><p>Here’s the uncomfortable truth: the moment data enters an AI inference pipeline, it is no longer “at rest.” It becomes input. Context. Training signal. A vector embedding that persists in ways no DLP policy anticipated. The model doesn’t steal your data. It <em>learns</em> your data — and that learning can propagate to edge nodes, get surfaced in unrelated queries, or be reconstructed from cached inference states that nobody thought to log.</p><p>This is what “data-in-use” risk actually means, and most enterprises have no framework for it. Encryption at rest doesn’t cover it. Tokenization doesn’t cover it. Anonymization barely slows it down when an agent has enough relational context to re-identify patterns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iQ4b2fvuoQ-p23WAwKqS-A.png" /></figure><p>The model isn’t breaking the rules. It’s operating exactly as designed — inside a policy framework that was designed for a world without autonomous inference.</p><h4>Are You Watching Everything Your AI Does?</h4><p>The industry answer to AI risk has been the Human-in-the-Loop. Keep a person in the chain. Make sure someone reviews the output before action is taken. Problem solved.</p><p>Except it isn’t. Not even close.</p><p>Here’s what Human-in-the-Loop actually looks like in production: a compliance officer reviewing 3% of flagged outputs on a Tuesday afternoon, using a UI dashboard that refreshes every four hours, approving or rejecting decisions they often lack the domain context to properly evaluate. The “loop” is a formality. The accountability is theatrical.</p><p>And in real-time AI deployments — financial risk scoring, supply chain routing, fraud detection, clinical triage support — there often is no loop. The velocity of inference outpaces the velocity of human review. The AI makes fifty decisions by the time a human has read the first summary.</p><blockquote><em>When your security perimeter ends at the AI layer, accountability collapses entirely into end-user judgment. That’s not governance. That’s hope.</em></blockquote><p>The more capable and autonomous your AI becomes, the less meaningful human review becomes — and the more essential automated, policy-enforced governance at every step of the decision chain actually is.</p><p>You can’t review your way to compliance. You have to architect your way to it.</p><h4>Accountable Decision Intelligence — What It Actually Requires</h4><p>If the leash analogy holds, then what you need isn’t a better vetting process at the door. You need something that travels with the AI everywhere it goes — enforcing boundaries continuously, logging everything, and never assuming good behavior is the same as governed behavior.</p><p>That framework has four non-negotiable layers:</p><ol><li><strong>Federated Execution:</strong> Models travel to the data. Data never travels to the model. Inference, RAG, and agent execution happen in-region, inside the governance boundary — not exported to centralized cloud endpoints. The answer to data sovereignty isn’t locking down egress. It’s ensuring the AI never requires egress in the first place.</li><li><strong>Zero Implicit Trust : </strong>No tool, agent, or workflow inherits permissions from the last one. Every interaction — human or model — is authenticated, authorized, and attribute-checked against live policy. Access isn’t granted at login. It’s enforced at every step, on every action, against every data asset touched.</li><li><strong>Decision Provenance: </strong>Every AI-assisted decision must have an auditable trail — not just that it happened, but why. What data was accessed? What model version? What policy context was active? What was the confidence? Compliance isn’t about knowing the outcome. It’s about being able to reconstruct the entire reasoning chain on demand.</li><li><strong>Runtime Policy Enforcement: </strong>Governance can’t be a pre-deployment checklist. Policies must be enforced in real time, at inference, across every agent action — automatically. Role-aware access controls that apply equally to humans, models, and agents. No exceptions for speed. No carve-outs for “low-risk” pipelines.</li></ol><h4>The Leash You Actually Need</h4><p>What enterprise AI actually requires is a control plane — something that sits between your agents and your data, enforcing policy at every interaction, logging every decision, and ensuring that the AI’s autonomy is always bounded by the enterprise’s rules. Not just at the door. Everywhere.</p><p>The most rigorous deployments we’re seeing today share a common architecture: a federated AI layer where models operate on data in-place rather than pulling data out of its governance boundary. Where the MCP (Model Context Protocol) layer doesn’t just connect agents to data — it governs how agents access data, invoke actions, and contribute to decisions, with security, policy, and accountability enforced at every step.</p><p>That means attribute-based authorization. Role-aware access controls that apply to agents the same way they apply to humans. Built-in audit logging of every agent interaction — not as an afterthought, but as infrastructure. And crucially: a decision provenance layer that captures not just <em>what</em> the AI decided, but the full context that drove it, reconstructable on demand for compliance, audit, or investigation.</p><p>This is what it looks like to actually hold the leash: not slowing your AI down, but ensuring that every step it takes is authorized, logged, and explainable. Speed and governance aren’t in tension. They’re both outcomes of getting the architecture right.</p><p>Platforms built around this philosophy — federated inference, zero implicit trust between agents and data, built-in decision provenance, and real-time policy enforcement across both human and AI actors — represent a fundamentally different bet than the “vet it at the door and hope” approach. They treat governance as infrastructure, not inspection.</p><p>Your AI probably has good nodes. Well-trained. Well-intentioned. Delivering real value.</p><p>Leash it anyway.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3fe714610bf0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Death of the ‘Approve’ Button]]></title>
            <link>https://medium.com/@poojapkamath/the-death-of-the-approve-button-e58e4e72ebda?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/e58e4e72ebda</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[human-in-the-loop-ai]]></category>
            <category><![CDATA[engineering]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 20:14:31 GMT</pubDate>
            <atom:updated>2026-04-01T20:14:31.721Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why active engagement is a safety requirement.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rTS2O0w2vPTrmJRt" /></figure><p>It is 11:45 PM. A wide-body freighter is sitting on a dark tarmac, loaded with millions of dollars in temperature-sensitive pharmaceuticals and high-risk lithium batteries. The AI’s load plan is ‘perfect’ — it’s optimized to save $12,000 in fuel. On the screen, a green button glows, pulsing with the word <strong>APPROVE</strong>.</p><p>In the next sixty seconds, the Loadmaster will either save the company a five-figure sum or accidentally ignite a mid-air chemical fire. The problem? The AI isn’t wrong about the math; it’s just blind to the context. This is the moment where ‘Human-in-the-Loop’ usually fails — because we’ve designed systems for speed, not for judgment.</p><p>The transition from a simple “Approve” button to a detailed <strong>Human-in-the-Loop (HITL)</strong> architecture isn’t just a technical upgrade — it’s a cultural shift. It moves the human operator from the role of a passive observer (who is often blamed when things go wrong) to an active <strong>Orchestrator</strong> (who is empowered to ensure things go right).</p><p>In high-consequence environments, the goal of Decision Intelligence is not to replace human judgment, but to <strong>dignify it.</strong> By providing evidence, allowing for “What-If” simulations, and capturing the nuance of every override, we create a system that doesn’t just work <em>for</em> us, but learns <em>from</em> us. We aren’t just building faster workflows; we’re building a more resilient, accountable, and transparent future for automated decision-making.</p><h3>The Four Pillars of Detailed HITL (Human In the Loop) Architecture</h3><p>Moving from a “rubber stamp” to an active orchestration model requires more than just a policy change; it requires a fundamental redesign of the digital workspace. Here is the four-stage framework for a true Decision Intelligence architecture:</p><h4>1. The “Graded Recommendations” Layer.</h4><p>Instead of providing one “best” answer, the AI must present a <strong>Decision Matrix</strong>. This shifts the human’s role from a passive “Yes/No” to an evaluator of trade-offs.</p><ul><li><strong>Confidence Scoring with Context:</strong> The AI must show its “work.” If it suggests Option A, it needs to display the underlying metrics — e.g., <em>92% regulatory compliance vs. 60% fuel efficiency.</em></li><li><strong>Risk Categorization:</strong> Recommendations are color-coded by their “Blast Radius.”</li><li><strong>Green:</strong> Routine/Safe.</li><li><strong>Yellow:</strong> Requires Senior Review (Policy boundary).</li><li><strong>Red:</strong> Policy Deviation/Safety Critical.</li><li><strong>Alternative Identification:</strong> The system should explicitly state its “runner-up” logic: <em>“I recommended Path A, but Path B was the secondary choice because it prioritizes speed over cost.”</em></li></ul><h4>2. Active Interrogation.</h4><p>A resilient HITL system allows the specialist to “stress test” the AI’s logic before the trigger is pulled.</p><ul><li><strong>The “What-If” Simulator:</strong> This is a sandbox within the workflow. The user can manually tweak variables — like a 2-hour crew delay or a sudden temperature spike — and watch the AI’s recommendation and risk scores recalculate in real-time.</li><li><strong>Mandatory Discrepancy Notes:</strong> If the AI’s suggestion conflicts with a hard-coded safety manual (like an FAA rule), the “Approve” button is locked. It only unlocks once the human provides a <strong>Mandatory Discrepancy Note</strong>, explaining why they are choosing to override a safety guardrail.</li></ul><h4>3. The Human-Centric Interface</h4><p>In technical operations, a chat box is a bottleneck. Detailed HITL utilizes <strong>Information Densification</strong> to speed up human recognition.</p><ul><li><strong>Side-by-Side Evidence Pinning:</strong> To prevent “tab-switching fatigue,” the AI’s recommendation should be pinned directly next to the source evidence — such as a technical log or a manual excerpt — allowing the human eye to verify the data instantly.</li><li><strong>The Visual Delta:</strong> Instead of a wall of text, the UI should highlight the <strong>Delta</strong> — the exact differences between the “Current State” and the “Proposed Future State.” This allows the human to spot anomalies in seconds rather than minutes.</li></ul><h4>4. The Recursive Feedback Loop.</h4><p>The “Intelligence” in DI comes from treating every human intervention as a high-value data point.</p><ul><li><strong>The “Edit-to-Learn” Log:</strong> If a specialist modifies the AI’s suggested parameters, that change is captured as a <strong>Correction Signal</strong> to refine future model outputs.</li><li><strong>Rationale Tags:</strong> For every override, the human provides a quick categorical tag (e.g., <em>“Weather context missing,” “Seniority exception”</em>). This builds an immutable audit trail for future training and regulatory compliance.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2GxKjmP5BqRomZk4rMcVRA.png" /><figcaption>Detailed HITL: Comparison Table</figcaption></figure><p>Moving from a ‘rubber stamp’ to a ‘detailed’ model requires more than just a change in policy; it requires a fundamental redesign of the workspace. We have to stop treating the human as a backstop for AI errors and start treating them as the primary investigator. This necessitates a structural blueprint that moves the operator through a logical sequence: from absorbing evidence to testing assumptions, and finally, to informed authorization.</p><h3>The Architectural Blueprint</h3><p>To visualize this, imagine the <strong>Human Intervention Node</strong> as the cockpit of the decision process.</p><h4>1. Ingest: The Evidence Review</h4><ul><li><strong>Action:</strong> The human reviews the “Work” (Logic + Sources).</li><li><strong>UI:</strong> Side-by-side comparison of the AI’s plan vs. the regulatory manual.</li></ul><h4>2. Simulation: The Stress Test</h4><ul><li><strong>Action:</strong> The human adjusts variables (time, cost, risk).</li><li><strong>UI:</strong> Real-time recalculation of risk scores based on “What-If” inputs.</li></ul><h4>3. Authorization: The Accountability Signature</h4><ul><li><strong>Action:</strong> Digital sign-off with a rationale tag.</li><li><strong>UI:</strong> A “Double-Sign” requirement for high-risk (Red-Level) decisions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*51XdKQXAsYtCeg6Eg6MK8w.png" /><figcaption>Detailed HITL</figcaption></figure><h3>Case Study: The Invisible Balancing Act</h3><h4>The Scenario: Flight 802 and the “Perfect” Load</h4><p>It is 11:45 PM at a global hub. A wide-body freighter is prepping for a cross-oceanic haul. The AI Load Planning system — a marvel of efficiency — calculates the optimal “trim.” To save $12,000 in fuel, it suggests a tight, rear-heavy distribution. On paper, the math is flawless.</p><p>But there’s a problem the AI “saw” but didn’t “understand.” The cargo includes five pallets of lithium-ion batteries and a pressurized container of live, temperature-sensitive pharmaceuticals.</p><p>In a <strong>Legacy AI</strong> world, the loadmaster sees a “Green Light” and hits <em>Execute</em>. In a <strong>Decision Intelligence</strong> world, the system behaves differently.</p><h4>1. The Evidence Ledger: Demanding the “Why”</h4><p>Instead of a simple loading chart, the DI system provides a <strong>Traceability Flow</strong>.</p><ul><li><strong>The Receipt:</strong> The system cites the <em>Aircraft Weight and Balance Manual (WBM)</em> Section 4.2.</li><li><strong>The Alert:</strong> A red flag appears. The AI’s suggested rear-heavy trim, while fuel-efficient, puts the lithium batteries too close to a heat-generating hydraulic line — a violation of the airline’s specific <strong>Standard Operating Procedures (SOPs)</strong> for hazardous materials.</li><li><strong>The Result:</strong> The system didn’t “smooth over” the conflict; it forced the discrepancy into the light.</li></ul><h4>2. The Specialist’s Shield: The Co-Pilot Model</h4><p>The system doesn’t just give one “correct” load plan. It presents the Loadmaster with a <strong>Tactical Menu</strong>:</p><ul><li><strong>Option A (Aggressive):</strong> Max fuel savings. (Flagged: High Risk due to HazMat proximity).</li><li><strong>Option B (Balanced):</strong> Moderate fuel savings. Batteries moved to a forward, cooler compartment.</li><li><strong>Option C (Conservative):</strong> Standard loading. Zero fuel optimization, maximum safety buffers.</li></ul><p>The Loadmaster, seeing the “Aggressive” option’s risk score of 85%, selects <strong>Option B</strong>.</p><h4>3. The Accountability Framework: The Human Signature</h4><p>Because this flight involves “Special Loads” (HazMat and Life Sciences), the <strong>Delegation of Authority</strong> shifts to <strong>Human-Led</strong>. The system prevents the final trim sheet from being sent to the cockpit until the Loadmaster provides a digital signature, confirming they have reviewed the “Contradiction Alert” regarding the battery placement.</p><h4>4. The Evolution Signal: The Loop</h4><p>The Loadmaster leaves a note in the system: <em>“Overrode Option A because seasonal ground temperatures exceed the AI’s current ambient air model for Tarmac Wait Time.”</em> This note isn’t lost in a database. It is captured as a <strong>Training Signal</strong>. The next time the AI plans a load in 90-degree weather, it will “remember” that tarmac heat is a contextual constraint it previously ignored.</p><h3>Why Traceability is Your Only Competitive Advantage</h3><p>If you can’t prove how an AI-assisted decision was made — what data informed it, which policies governed it, and who ultimately stood behind it — <strong>you are exposed.</strong> In a high-stakes environment, “the algorithm suggested it” is not a defense; it is a confession of a lack of control.</p><p>However, there is a second, quieter problem that most organizations miss: <strong>If you can’t trace the “why,” you are also flying blind on whether your decisions are actually good.</strong></p><blockquote><strong>The “Approve” button is dead. Long live the Pilot-in-Command.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e58e4e72ebda" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decision Intelligence : From AI output to accountable decisions.]]></title>
            <link>https://medium.com/@poojapkamath/decision-intelligence-from-ai-output-to-accountable-decisions-a09c4966df39?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/a09c4966df39</guid>
            <category><![CDATA[human-in-the-loop]]></category>
            <category><![CDATA[decision-intelligence]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[audit-trail]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 15:31:37 GMT</pubDate>
            <atom:updated>2026-04-01T20:16:46.447Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8G3ErflOQrLjD8Sa" /></figure><h3>The $500 Million Math Error</h3><p>The glowing green light on the dashboard didn’t flicker. It beamed.</p><p>In the high-tech war rooms of Seattle, Zillow’s proprietary AI — a digital oracle fed on a decade of housing data — had reached a verdict. The command was simple, cold, and incredibly ambitious: <strong>Buy.</strong> The models were masterpieces of technical sophistication. They looked at the post-pandemic horizon and saw nothing but clear skies and climbing prices. Armed with these “perfect” outputs, Zillow’s <em>Offers</em> division moved into the streets like a machine. They weren’t just participating in the real estate market; they were outbidding humans with the soulless precision of an algorithm that believed it couldn’t lose.</p><p>For a moment, they were the masters of the neighborhood. Until the world changed, and the algorithm didn’t notice.</p><p>By November 2021, the dream didn’t just fade — it hemorrhaged. The “Buy” signal had led them over a cliff. In a series of brutal emergency meetings, the reality set in: the division was dead. The fallout was cinematic in its devastation — 2,000 employees shown the door and a staggering <strong>$500 million</strong> in real estate inventory vaporized into a write-off.</p><p>When the dust settled, the industry “coroners” performed their post-mortem. They looked for a glitch. They searched for a broken line of code or a corrupted dataset.</p><p>They found nothing.</p><p>The AI had done exactly what it was told to do. The failure wasn’t in the <em>intelligence</em> — it was in the <em>decision</em>. Zillow had the world’s best compass, but no one was looking at the map. They lacked the layer of accountability that asks the only question that matters when the stakes are half a billion dollars:</p><p><em>“What happens if the model is wrong, and who is standing at the wheel when it is?”</em></p><p>Zillow’s story isn’t a tech failure; it’s a <strong>Decision Intelligence</strong> ghost story. It’s the moment we realized that an AI “output” is just a suggestion — and without a framework for accountability, a suggestion can be a catastrophe.</p><h3>The Three Pillars of Decision Intelligence (DI)</h3><p>If we treat AI as the engine, Decision Intelligence is the steering wheel, the brakes, and the GPS combined. To move from “blindly trusting models” to “engineering outcomes,” you need a framework built on three non-negotiable pillars:</p><h3>1. Data Contextualization (The “Why”)</h3><p>AI models are brilliant at finding correlations but terrible at understanding <strong>why</strong> they exist.</p><ul><li><strong>The Gap:</strong> A model might see a spike in hospital admissions and recommend buying more beds.</li><li><strong>The Intelligence:</strong> A human understands there’s a blizzard outside and the beds aren’t the problem — transportation is.</li><li><strong>The Goal:</strong> DI works only when technical outputs are grounded in context and constraints.</li></ul><p>In Zillow’s case contextualization would have flagged that “Predicted Profit” is irrelevant if “Time-to-Renovate” is doubling due to labor shortages.</p><h3>2. Decision Modeling (The “How”)</h3><p>Most companies treat decisions as a single point in time. In reality, a decision is a <strong>process</strong>.</p><ul><li><strong>The Shift:</strong> Instead of asking “What is the prediction?”, DI asks “What is the desired outcome, and what actions lead there?”</li><li><strong>The Action:</strong> This involves mapping out the “Decision Flow” — identifying every human and machine touchpoint to ensure the path to the finish line is visible and repeatable.</li></ul><p>In Zillow’s case, A decision model would have set <strong>Thresholds of Uncertainty</strong>. For example: <em>“If our inventory exceeds X and our average days-on-market exceeds Y, the AI’s buying power is automatically throttled by 50%.”</em></p><h3>3. The Accountability Loop (The “So What?”)</h3><p>This is where Zillow failed. There was no feedback loop to catch the drift between the model’s “truth” and reality.</p><ul><li><strong>The Mechanism:</strong> Every AI-backed decision must have a “Human-in-the-loop” or a “Guardrail” system. If the model suggests an action that exceeds a certain risk threshold, the system must trigger a manual review.</li><li><strong>The Result:</strong> You create a clear audit trail. When a decision is made, you know exactly which data influenced it, which model suggested it, and which human signed off on it.</li></ul><p>In Zillow’s case, An Accountability Loop would have required a regional manager to “Sign Off” on any purchase that deviated from historical price averages by more than 10%. It puts a human hand on the emergency brake before the train leaves the tracks.</p><blockquote>Beyond the Chatbot: Engineering Defensible Action</blockquote><h4>The Architecture of Accountability</h4><p>Moving from “AI output” to “Decision Intelligence” (DI) is the shift from a probabilistic guess to a structured, defensible action. In a professional landscape where stakes are high — whether in global logistics or technical operations — an AI’s “chat” response isn’t a decision; it’s just data.</p><p>To bridge that gap, the architecture must move toward Accountable Decision Systems.</p><h4>1. The Evidence Audit Trail (The “Receipt”)</h4><p>A standalone AI prediction is a “black box” — a conclusion without a cause. In a high-stakes environment, an output is worthless if it isn’t defensible. True Decision Intelligence requires more than just “Explainable AI” (XAI); it requires <strong>Data Lineage</strong> — a transparent, immutable record of exactly how a conclusion was reached.</p><p>To build an audit trail that holds up under scrutiny, three forensic layers are required:</p><ol><li><strong>Ground-Truth Grounding (Fact-Checking):</strong> We must move past the “trust me” era of AI. Every recommendation must come with a verifiable bibliography — citing specific technical manuals, real-time telemetry, or historical records. If the AI suggests a course of action, it must show the “receipts” that justify the logic.</li><li><strong>Friction by Design (Contradiction Alerts):</strong> Large Language Models (LLMs) are designed to be helpful, which often leads to “smoothing over” inconvenient truths. A DI system must do the opposite: if a suggestion conflicts with established Standard Operating Procedures (SOPs) or regulatory constraints, the system should flag the discrepancy as a high-priority alert rather than trying to reconcile it.</li><li><strong>The Temporal Snapshot (Version Control):</strong> Accountability requires a “time machine” for data. Just as you track code changes, DI requires tracking the exact model version and the specific data snapshot used at the moment of the decision. When things go wrong, you need to know exactly what the machine “knew” at the time of the trigger.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qtz8DDDDFuT6Z6Sks0-Hsg.jpeg" /></figure><h4>2. Human-in-the-Loop (HITL) Workflows</h4><p>Automation is a light switch — it is either on or off, Go or No-Go. But high-stakes operations don’t live in a binary world; they live in the “gray.” True Decision Intelligence isn’t a replacement for human expertise; it is a force multiplier for it. To prevent the next Zillow-scale collapse, the specialist must remain the final arbiter through three tactical layers:</p><ol><li><strong>The Tactical Menu (The “Co-Pilot” Model)</strong></li></ol><p>In a flawed system, the AI gives an order. In a DI system, the AI provides a <strong>spectrum</strong>. Instead of a single “best” path, the specialist is presented with three graded strategies — <em>Conservative, Balanced, and Aggressive</em> — each accompanied by a raw risk-probability score. This transforms the AI from a “black box” oracle into a sophisticated advisor, leaving the final strategic choice to the person who bears the consequences.</p><p>2.<strong> The Moral Brake (Active Intervention)</strong></p><p>Accountability cannot be automated. In high-risk environments, “The algorithm told me to” is a catastrophic defense. DI workflows must implement a <strong>Risk Threshold</strong>: any action that carries a significant “blast radius” requires a physical or digital human signature to execute. This ensures that a human mind — capable of interpreting context that code might miss — is always standing at the wheel during a storm.</p><p><strong>3. The Evolution Signal (Feedback Loops)</strong></p><p>The most valuable data point in your organization isn’t an AI success; it’s a <strong>human override</strong>. When a specialist rejects an AI recommendation, it shouldn’t be a silent click. That rejection must be captured as a high-priority training signal. These “human-led course corrections” are the only way to refine a model’s understanding of edge cases, turning today’s human intuition into tomorrow’s systemic intelligence</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ieELa4J39qUQ0mVlRI6YIg.jpeg" /></figure><h4>3. Beyond Chat: Functional Interfaces</h4><p>A text stream, while great for processing information, is a poor vehicle for a $500 million decision. For experts navigating high-stakes technical or corporate environments, a chat interface is often entirely unsuitable, stripping away the critical spatial context and rich data visualization that human reasoning requires.</p><h4>4. The Accountability Framework</h4><p>Accountability isn’t a checkbox; it’s a spectrum. To prevent a systemic collapse, organizations must treat AI like an aircraft’s flight system — knowing exactly when to engage the autopilot and when the Captain must take the stick.</p><ol><li><strong>The Autopilot (Machine-Led):</strong> Used for the mundane and the repetitive. In logistics, this is re-routing a minor delay. It happens in milliseconds, optimized for efficiency where the cost of error is negligible.</li><li><strong>The Co-Pilot (Machine-Assisted):</strong> Used for complex navigation. The AI maps the route and warns of turbulence, but the human pilot confirms the course change. This is the “Collaborative” layer where nuance meets math.</li><li><strong>The Captain (Human-Led):</strong> Reserved for the emergency landing. In safety-critical or high-capital environments, the AI provides the telemetry, but the human makes the judgment call. This is where <strong>Decision Intelligence</strong> saves lives — and balance sheets.</li></ol><h3>The DI Lifecycle: From Signal to Signature</h3><p>To move beyond the “Black Box,” an organization needs more than just a model; it needs a functional workflow that mirrors the way experts actually work. This is the lifecycle of a defensible decision:</p><h3>Phase 1: Monitor &amp; Explore (The Radar)</h3><p>Before a decision is made, there must be a state of awareness. A DI platform doesn’t just show data; it monitors for <strong>Signals</strong>.</p><ul><li><strong>Contextual Alerts:</strong> Configure signals based on your specific operational data to highlight issues and incidents before they escalate.</li><li><strong>Active Investigation:</strong> Move seamlessly from a “notification” to a deep-dive investigation within a dedicated data space, exploring the full breadth of available telemetry.</li></ul><h3>Phase 2: Investigate &amp; Analyze (The Engine)</h3><p>Once an incident is identified, the “Co-Pilot” engages. This is where AI moves from a static output to a dynamic collaborator.</p><ul><li><strong>Pluggable Intelligence:</strong> Engage with a natural language AI assistant. Whether utilizing a specific LLM or a custom-tuned model, the interface should allow for fluid, iterative analysis.</li><li><strong>Strategic Guidance:</strong> The system shouldn’t just “chat”; it should provide <strong>Recommended Queries</strong> and <strong>Resolution Paths</strong>, guiding the specialist through the data rather than just summarizing it.</li></ul><h3>Phase 3: Freeze &amp; Attest (The Ledger)</h3><p>This is the most critical stage — where the “Decision” becomes “Accountable.”</p><ul><li><strong>The Decision Manifest:</strong> Users “pin” relevant evidence (telemetry, manuals, chat logs) to create a definitive manifest of the decision-making process.</li><li><strong>The Immutable Record:</strong> Once approved, the entire package — Evidence + AI Logic + Human Notes — is <strong>attested and recorded</strong>.</li><li><strong>The Forensic Replay:</strong> Accountability means a reviewer can “replay” the decision record months later, seeing exactly what the specialist saw and why they pulled the trigger.</li></ul><h3>Final Verdict: The New Standard of Leadership</h3><p>We are exiting the era of “The Black Box.”</p><p>The Zillow disaster taught us that technical brilliance is a liability without a decision framework. Whether you are buying houses, planning flight paths, or managing a global supply chain, the goal of AI integration should never be “Autonomy.” The goal is <strong>Agency</strong>.</p><p><strong>Decision Intelligence is the bridge between a probabilistic guess and a defensible action.</strong> It transforms the specialist from a passive observer of an algorithm into an empowered architect of outcomes.</p><h3>Checklist for Leaders:</h3><ol><li><strong>Audit your interfaces:</strong> Is your AI “chatting” at you, or is it showing you the evidence?</li><li><strong>Define your DoA (Delegation of Authority):</strong> Do you know which decisions are machine-led and which require a “Moral Brake”?</li><li><strong>Capture the Overrides:</strong> Are you treating human intuition as “noise” or as the most valuable training data you own?</li><li><strong>Dashboards: </strong>Can you see at a glance your Active signals and history.</li><li><strong>Evaluate the Lifecycle:</strong> Does your current tooling allow you to <strong>Freeze and Attest</strong> a decision, or are your AI’s reasons vanishing the moment you close the chat window?</li></ol><p>The future of industry won’t be won by the company with the smartest AI. It will be won by the company that makes the best decisions —puts the experience of its humans at the center of the machine.</p><p><em>Interested in a deep dive? Part 2 is here: </em><a href="http://hitl/">https://medium.com/@poojapkamath/the-death-of-the-approve-button-e58e4e72ebda</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a09c4966df39" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Structuring User Management API with OAS in MuleSoft]]></title>
            <link>https://medium.com/@poojapkamath/structuring-user-management-api-with-oas-in-mulesoft-3100eb6445ef?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/3100eb6445ef</guid>
            <category><![CDATA[openapi-specification]]></category>
            <category><![CDATA[mulesoft-training]]></category>
            <category><![CDATA[anypoint-platform]]></category>
            <category><![CDATA[mulesoft-integration]]></category>
            <category><![CDATA[api-design]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Mon, 12 May 2025 22:45:30 GMT</pubDate>
            <atom:updated>2025-05-12T23:06:24.862Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Life’s unpredictable. Your API shouldn’t be.</em></h4><p>One day, while my brain was clearly stuck in debug mode, my daughter walked in and started firing off questions about the weekend party I was hosting.</p><blockquote><em>“Who’s coming? What time? What am I responsible for? Are there snacks? Can we watch TV?”</em></blockquote><p>I had no answers, I was the flaky backend, and my daughter was trying to consume an undocumented API without clear instructions.</p><h3>Why does structure matter?</h3><p>The structure for an API is essential, so API users, like my daughter, are not left guessing what endpoints exist, what data they need, and what responses to expect.</p><p>That’s where OpenAPI Specification (OAS) comes in. It’s like giving your API a user manual, a contract, and a personality — all in one YAML file.</p><h3>Structuring an OAS API in the Anypoint Platform</h3><p>1. Go to <a href="https://anypoint.mulesoft.com/">https://anypoint.mulesoft.com</a></p><p>2. Log in with your credentials.</p><h4>Step 2: Create a new API specification</h4><ol><li>Navigate to Design Center → Create new API specification</li><li>Choose OpenAPI Specification (OAS) 3.0</li><li>Name your API: User Management API</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n5ZYM61fy6H9hbY7-lxbag.gif" /></figure><h4>Step 3: Define the API Structure</h4><ol><li>Set up basic metadata.</li></ol><pre>openapi: 3.0.0<br>info:<br>  title: User Management System API<br>  description: API for managing user data in a backend system<br>  version: 1.0.0<br>servers:<br>  - url: https://api.company.com/system/users<br>    description: Production server</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*es12vUvGVWpxGBxVObiuxw.gif" /></figure><p>2. Define basic <strong>paths </strong>and <strong>operations</strong>, for example:</p><pre><br>paths:<br>  /users:<br>    get:<br>      summary: Get all users<br>      description: Retrieves a list of all users in the system<br>      responses:<br>        &quot;200&quot;:<br>          description: A list of users<br>          content:<br>            application/json:<br>              example:<br>                users:<br>                  - id: 1<br>                    name: User1<br>                    email: user1@example.com<br>                  - id: 2<br>                    name: User2<br>                    email: user2@example.com</pre><p>3. Try it to check everything looks okay.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B7bUB3ZcD7NrSaljEbk8Yw.gif" /><figcaption>try-the-api</figcaption></figure><h4>Step 4: <strong>Add Authentication &amp; Security</strong></h4><ol><li>You can include <strong>OAuth 2.0 or API Key security</strong>:</li></ol><pre>components:<br>  securitySchemes:<br>    ApiKeyAuth:<br>      type: apiKey<br>      in: header<br>      name: Authorization<br>security:<br>  - ApiKeyAuth: []</pre><p><strong>Step 4: Publish and Deploy</strong></p><ol><li>Publish API to <strong>Exchange</strong> or deploy via API Manager.</li><li>You can use <strong>Mocking Service</strong> in MuleSoft to test responses.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UVxr0eTwayuy4ivazz4l8w.gif" /><figcaption>publish-the-api</figcaption></figure><p>And there you have it — your OpenAPI Specification!</p><p>Happy Coding — May your requests always return <strong>200 OK</strong>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3100eb6445ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top Reasons to Enroll in Anypoint Platform Development Fundamentals Training.]]></title>
            <link>https://medium.com/another-integration-blog/top-reasons-to-enroll-in-anypoint-platform-development-fundamentals-training-1991ab913033?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/1991ab913033</guid>
            <category><![CDATA[mulesoft-certification]]></category>
            <category><![CDATA[mulesoft-training]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Mon, 10 Jul 2023 07:13:34 GMT</pubDate>
            <atom:updated>2023-07-10T07:13:34.123Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*m04Q06HnZIZuXwJe" /><figcaption>Photo by <a href="https://unsplash.com/@andrewtneel?utm_source=medium&amp;utm_medium=referral">Andrew Neel</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>As an experienced MuleSoft professional with numerous end-to-end project implementations under my belt, I was pleasantly surprised at how beneficial I found the Anypoint Platform Development Fundamentals course, which I recently completed.</p><p>Do you know how it feels when you pick up a book you have already read multiple times? This felt a lot like that, even when you are familiar with the topic, revisiting it helps reinforce your existing knowledge and provide a refresher on key concepts that you may have forgotten or overlooked. it provides a new perspective or fresh insights that can give you a deeper understanding of the MuleSoft fundamentals. It can also help identify gaps in your knowledge, allowing you to identify areas where you may need to focus more attention or seek out additional resources. In my case, it helped me get familiar with new developments or updates that I was previously unaware of.</p><p>On the Trailhead Academy website, under course it says..</p><p><strong><em>“This course is for developers and architects who want to get hands-on experience using the Anypoint Platform to build APIs and integrations. It prepares you for the MuleSoft Certified Developer — Level 1 exam and includes two attempts.”</em></strong></p><h4>What’s in it for New MuleSoft Developer?</h4><p>This comprehensive course provides structured learning that’s designed specifically for newcomers to the platform. The course itself is structured to be ramp up new developers quickly, that being said, it’s imperative for everyone taking the course to know REST and basic programming concepts. The course focuses on Core MuleSoft and is not a good fit for application testers, DevOps engineers, or Analysts. If you take the course, you will benefit from logical concepts and hands-on exercises that reinforce everything you learn.</p><p>In this course, every instructor is a highly skilled subject matter expert with extensive practical experience working with MuleSoft. Their expertise will be invaluable in guiding you to effectively and efficiently utilize the technology, enabling you to achieve mastery quickly. Additionally, instructors will provide valuable feedback on assignments and projects, and I strongly recommend taking advantage of this opportunity. Often, I have observed students neglecting assignments and postponing them for later, but the course fee you pay includes instructor feedback and assistance. This feature sets instructor-led training apart from self-paced learning since you have immediate access to instructor feedback.</p><p>Completing course exercises and assignments will help you quickly identify areas where you need to improve and gain a deeper understanding of the core concepts in MuleSoft.</p><p>Tips for new developers:</p><ol><li>Install MuleSoft studio before coming to the class. Follow this to learn how <a href="https://docs.mulesoft.com/studio/7.15/to-download-and-install-studio-wx">https://docs.mulesoft.com/studio/7.15/to-download-and-install-studio-wx</a></li><li>Create an Anypoint Platform account for your training. This is free and is usually active for 30 days.</li><li>If possible complete the self-paced fundamentals training found here <a href="https://training.mulesoft.com/course/development-fundamentals-mule4-bootcamp">https://training.mulesoft.com/course/development-fundamentals-mule4-bootcamp</a>.</li></ol><p><strong>What’s in it for Seasoned MuleSoft Developers?</strong></p><p>Even experienced professionals can benefit from taking courses to expand their knowledge and skills in specific areas. These courses can help you become more valuable to your organization and more competitive in the job market. And, with a dedicated instructor, you’ll feel accountable and motivated to stay on track with your learning, pushing you to achieve your goals faster than you ever thought possible.</p><p>Tips for Seasoned MuleSoft Developers:</p><ol><li>Set up a dedicated workspace specifically for your training purposes to avoid any potential issues arising from project naming conflicts or configuration problems.</li><li>Download all student files ahead of time.</li><li>Stay on top of your assignments.</li><li>Create a comprehensive list of things/topics that you wish to prioritize, as having a well-structured list will aid in maintaining focus and maximizing the benefits of your training.</li><li>Make sure to take notes, as many of these topics will already be familiar to you, giving you a special advantage in creating a valuable reference for the future.</li></ol><h4>After Training</h4><p>Training is complete, now what? Well, Salesforce offers a free certification voucher which has two exam attempts, the exam details can be found here <a href="https://training.mulesoft.com/certification/developer-mule4-level1">https://training.mulesoft.com/certification/developer-mule4-level1</a>. Take it while the concepts are still fresh in your mind. Certification exams offer numerous benefits that can positively impact your career trajectory, professional brand, and personal development.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*E-hObKCt_FDHdAud.png" /><figcaption>Sample certification badge</figcaption></figure><p>If you take the course or the certification do tag me on Twitter @poojapkamath and let me know how it went. I am rooting for you!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1991ab913033" width="1" height="1" alt=""><hr><p><a href="https://medium.com/another-integration-blog/top-reasons-to-enroll-in-anypoint-platform-development-fundamentals-training-1991ab913033">Top Reasons to Enroll in Anypoint Platform Development Fundamentals Training.</a> was originally published in <a href="https://medium.com/another-integration-blog">Another Integration Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Discover the MuleSoft Integration Associate Certification: MuleSoft’s newest Credential]]></title>
            <link>https://medium.com/another-integration-blog/discover-the-mulesoft-integration-associate-certification-mulesofts-newest-credential-eb3d0bba4242?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb3d0bba4242</guid>
            <category><![CDATA[certification]]></category>
            <category><![CDATA[mulesoft]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Wed, 30 Nov 2022 22:57:24 GMT</pubDate>
            <atom:updated>2022-12-01T01:11:05.521Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why you should pursue the entry-level <a href="https://training.mulesoft.com/certification/associate-integration">MuleSoft Integration Associate Certification</a> and how to get started today.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*2wm0YtCS9rEZ-3mKbcH21g.jpeg" /></figure><p>As we emerge from a global pandemic and are trying to stave off the recessionary financial impact, we need to do what is within our power to shift out of survival mode. To improve lives and ensure access to opportunities, we must take a proactive and strategic approach toward upskilling and reskilling. The key to building a resilient workforce is facing the digital skill crisis head-on.</p><p>In December 2020, a MuleSoft-commissioned Forrester Total Economic Impact™ (TEI) study examined the potential return on investment (ROI) of participating in MuleSoft’s Training and Certification program*. The study found 3x increased job opportunities for individuals with a MuleSoft Certification and 25% time savings on significant projects. MuleSoft Certifications are currently available for key Developer and Architect roles, and with the introduction of the new MuleSoft Integration Associate Certification, they’ve expanded to cater to a broader audience.</p><p>During my mentoring journey, I have always been asked, if there was an alternative to MCD certification. Mostly, this was asked by people who were new to the development landscape or performing a non-coding IT role. Not everyone on an IT project is a developer, we need analysts, testers as well as project management folks to successfully drive a software initiative. If it’s a MuleSoft project, the developers, testers, and architects are usually familiar with MuleSoft, and the rest of the project team supports it from a non-mule perspective. Wouldn’t it be great if there was a way to ramp up the other participants in the project and arm them with the foundational knowledge required for the successful implementation of integration projects?</p><p><strong>A certification option for those new to learning MuleSoft</strong><br>The <a href="https://training.mulesoft.com/certification/associate-integration">MuleSoft Integration Associate Certification</a> is an entry-level credential designed for those who are new to learning MuleSoft (0–6 months of MuleSoft experience). The exam covers the foundational knowledge required for integration projects, with a wide breadth of foundational integration topics covered.<br>You’ll learn how to:</p><ol><li>Identify a typical integration project’s roles, responsibilities, and lifecycle phases.</li><li>Recognize and interpret essential integration concepts and terminology used by MuleSoft architects and developers</li><li>Recognize common integration problems, deconstruct them into their fundamental integration use cases, and identify the appropriate technologies to solve them.</li><li>Explain the common technical complexities that are central to integration development.</li><li>Describe the components and benefits of the Anypoint Platform for system integration and API management.</li></ol><p><strong>Who should pursue the MuleSoft Integration Associate Certification?</strong><br>This certification is for anyone interested or involved in integration projects. When working on a MuleSoft project, it is imperative to understand core integration concepts for successful implementations. It’s especially well-suited for team members that will work with MuleSoft developers and architects, including project managers, integration product owners, IT professionals, IT operations, and university students.</p><h3>Get started today</h3><h4>Prepare for the exam by:</h4><ul><li>Completing the <a href="https://trailhead.salesforce.com/users/mulesoft-trailhead/trailmixes/mulesoft-certified-integration-associate-credential">Trailmix</a>*: includes a list of tasks, links to Trailhead modules, documentation, whitepapers, and blog posts</li><li>Completing the <a href="https://training.mulesoft.com/certification/associate-integration/practice-exam">Practice exam</a>: is the same format and length as the exam with questions that are the same style and degree of difficulty as the exam</li><li>Reviewing the exam information and signing up on the <a href="https://training.mulesoft.com/certification/associate-integration">MuleSoft Certified Integration Associate page</a>.</li></ul><p><em>Note that a new MuleSoft training course supporting this exam will be available in 2023. Don’t forget to share your shiny new certification on socials with #MuleSoftTraining and #MuleSoftCommunity.</em></p><p>Happy learning!</p><p><a href="https://training.mulesoft.com/certification/associate-integration"><strong>MuleSoft Certified Integration Associate</strong></a><strong>: </strong>Validate that you can work as an informed member of a MuleSoft project team, with knowledge of core integration and API–led connectivity. (24 kB)</p><p><a href="https://www.mulesoft.com/lp/reports/developer-training-certification"><strong>The Total Economic Impact™ Of MuleSoft Training and Certification, a commissioned study conducted by Forrester Consulting, December 2020.</strong></a></p><h4>This Blog was Co-Authored by:</h4><ol><li><a href="https://medium.com/@nicole.isaacson">Nicole Issacson</a></li></ol><p>Nicole is a creative and strategic problem solver who’s always looking for ways to improve efficiency. Nicole has expertise in propelling value while engaging businesses and consumers more effectively. Nicole is passionate about go-to-market strategy, user-centered design, and telling a great story. Nicole is a Tech Marketer at <a href="https://medium.com/u/1d08fbe24b8b">MuleSoft</a>. Ask her about her team’s awesome work to upskill students on the integration platform.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb3d0bba4242" width="1" height="1" alt=""><hr><p><a href="https://medium.com/another-integration-blog/discover-the-mulesoft-integration-associate-certification-mulesofts-newest-credential-eb3d0bba4242">Discover the MuleSoft Integration Associate Certification: MuleSoft’s newest Credential</a> was originally published in <a href="https://medium.com/another-integration-blog">Another Integration Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Piloting Mule Apps from POC to Production on CloudHub]]></title>
            <link>https://medium.com/another-integration-blog/piloting-mule-apps-from-poc-to-production-on-cloudhub-62b2a2d4cb4f?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/62b2a2d4cb4f</guid>
            <category><![CDATA[production]]></category>
            <category><![CDATA[cloudhub]]></category>
            <category><![CDATA[apps]]></category>
            <category><![CDATA[integration]]></category>
            <category><![CDATA[mulesoft]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Mon, 26 Sep 2022 17:23:42 GMT</pubDate>
            <atom:updated>2022-09-26T17:23:42.524Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7UeGIuZEDJpgMs0XalaEKQ.jpeg" /><figcaption>Take-off with Mule</figcaption></figure><p>Your POC was a success, now what?</p><p><strong><em>Let’s deploy to production!</em></strong></p><p><strong><em>Deploy to production? Wait, WHATTTT?</em></strong></p><p><strong><em>This is usually when I realize, they didn’t listen to a single thing I said!</em></strong></p><p>But, is it that complicated? How do organizations plan to transition from POC to production? Is there an in-between? Why is it important to have the right strategy to ensure a successful transition? Depending on the technology used, scaling, security, governance, organizational enablement, and change management are usually needed.</p><p>MuleSoft recommends the 3-pillar approach for successful implementation:</p><ol><li><strong>Business outcomes<em>: </em></strong>align stakeholders and define outcomes with clear KPIs</li><li><strong>Organizational enablement<em>: </em></strong>ensure your organization is ready to use and adopt the Anypoint Platform</li><li><strong>Technology delivery:<em> </em></strong>get up and running with Anypoint Platform and start building APIs and integrations</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UufUdWyKLRpGe8SWKaJcJw.png" /><figcaption>MuleSoft recommends a 3-pillar approach for successful implementation</figcaption></figure><p>A POC is a Proof of Concept, an exercise to prove feasibility. With a POC, you are trying to understand if the tool will help meet your needs and whether the solutions are feasible.</p><p>MuleSoft prides itself on being a rapid development and deployment tool. In fact, products and projects developed using MuleSoft are 3x times more likely to be on an accelerated schedule compared to other development tools. Although software projects are rarely just about development and deployment, they usually involve planning, designing, end-user testing, and training.</p><p>With that in mind, the pillars are further broken into six paths for users to follow: business outcomes, Anypoint Platform, projects, Center for Enablement (C4E), internal support, and training.</p><p>The blueprint that summarizes tasks across these 6 paths serves as a template to track the success and progression of each desired business outcome. The blueprint is very comprehensive and customers usually only need a subset of these tasks at any given time. In the next section, I will highlight the important tasks that should be taken when accelerating from POC to production.</p><p>At minimum, to be in production, you will need to convert the POC to a pilot and then move it to production. For a POC, you may not need to address organization enablement; however, for pilot and production, all 3 pillars will be essential for successful implementation.</p><h3><strong>The In-between</strong></h3><p>This phase could take anywhere from weeks to months, depending on the size of the deployment. It is called the “pilot” phase. When you are piloting a MuleSoft project it usually means you have to do the following:</p><ul><li><strong>Refreshing success plan</strong>: if you completed a successful POC, then you have a plan, a definition of the outcome, and a way to measure that outcome. To move to pilot, you need to update your success plan to align with your new objectives.</li><li><strong>Updating the design</strong>: a detailed architecture design based on an assessment of business and technical requirements will be necessary. Along with this, you should also identify and define KPIs and align stakeholders.</li><li><strong>Implementation</strong>: build and configure the environment on CloudHub. You most likely created the POC in a development environment. For a pilot, setting up an environment like production is important to ensure effective user testing and to gather iterative feedback to optimize the application.</li><li><strong>Documentation or training for pilot users</strong>: formulate a training plan to educate end users.</li></ul><p>These activities can happen iteratively, in federated teams, to ensure timelines are managed appropriately. Most of these activities are independent of each other.</p><p>The other things you could do at this stage include:</p><p><strong>Define a support model</strong>: onboarding activities should be complete by now and defining a support model would be the next logical step. This will help identify training requirements for your support staff.</p><p><strong>Onboard MuleSoft teams: </strong>if the end product is going to be significantly larger than the POC, you may need additional resources to realize project goals.</p><p><strong>Establish C4E</strong>: some companies like to wait to form a C4E until they have a few projects in production; however, I recommend forming a C4E as early as possible.</p><p><strong>C4E bonus items: </strong>setting up the development environment, creating reusable libraries, establishing DevOps (if you do not already have one), or aligning the project to existing DevOps can also be part of your pilot initiative.</p><p>Since the infrastructure is owned by MuleSoft, it is robust and has mechanisms in place to help scale up or down with relative ease. Therefore, testing to validate failover, high availability, and scalability be simple and quick.</p><p>A <a href="https://developer.mulesoft.com/guides/quick-start/developing-your-first-mule-application">Mule application</a> is developed as a set of flows and each flow is a chain of event processors. An event processor processes data passing through the flow with logic from one of the said domains. Mule 4 uses reactive programming to facilitate the non-blocking execution of the event processors. This has a significant impact on a Mule application’s ability to scale the number of events it can handle concurrently. We will talk about scalability in detail later.</p><p>As the pilot progresses and is deemed successful, the environment can be transferred to production and implemented across the entire end-user community.</p><h3>Moving to Production</h3><p>The steps to production vary by organization. Before production is reached, organizations typically complete the following steps:</p><ul><li>Training for end-users and help desk support</li><li>Formal change management</li><li>Business acceptance</li><li>Sign-off from the business</li></ul><p>But, how do you get there?</p><p>The most challenging part of change is a coherent measurement, and in particular, the measurement and account for consumption. The model must be continually refined for effectiveness. With that in mind, organizations will have to :</p><ul><li>Refresh the success plan</li><li>Update design with volume in mind</li><li>Train/inform users</li><li>Pursue logical change management</li><li>Document all the efforts</li></ul><h3>The Elephant in the Room: Scalability</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j09X1E6epSFHXK8aKQ2KnQ.jpeg" /><figcaption><strong>Provisioning/scaling is fully managed in CloudHub 2.0</strong></figcaption></figure><p>Scalability is one of the biggest reasons why projects fail in production. The goal for any organization is to create a leaner product that can easily scale up and down.</p><p>Mule is highly scalable both horizontally and vertically. Automatic tuning of deployed Mule applications allows them to execute in a non-blocking fashion that exploits all resources available to the Mule runtime. Threads are assigned to the application in a way that maximizes concurrency.</p><h4>Replica Scale-out</h4><p>CloudHub 2.0 enables you to select more vCores for your application, providing horizontal scalability. This gives you the flexibility to scale up your application to handle higher loads (or scale down during low-load periods) at any time. This allows for greater control over capacity provisioning. Depending on your subscription, you can deploy your application with up to 8 replicas. To learn more, see <a href="https://docs.mulesoft.com/cloudhub-2/ch2-architecture#cloudhub-2-replicas">CloudHub 2.0 Replicas</a>.</p><h3><strong>Survival Tips</strong></h3><p>The journey from POC to production is not a linear progression. Moving applications to production is a multi-dimensional feat that requires planning and effort. The key to success is a continuous approach to design, documentation, and implementation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62b2a2d4cb4f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/another-integration-blog/piloting-mule-apps-from-poc-to-production-on-cloudhub-62b2a2d4cb4f">Piloting Mule Apps from POC to Production on CloudHub</a> was originally published in <a href="https://medium.com/another-integration-blog">Another Integration Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pizza, Wine, and Great Company!]]></title>
            <link>https://medium.com/another-integration-blog/pizza-wine-and-great-company-339818023a1?source=rss-4b73901802bb------2</link>
            <guid isPermaLink="false">https://medium.com/p/339818023a1</guid>
            <category><![CDATA[salesforce]]></category>
            <category><![CDATA[events]]></category>
            <category><![CDATA[community]]></category>
            <category><![CDATA[mulesoft]]></category>
            <dc:creator><![CDATA[Pooja Kamath]]></dc:creator>
            <pubDate>Mon, 22 Aug 2022 17:42:04 GMT</pubDate>
            <atom:updated>2022-08-22T17:42:04.275Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Disclaimer: No animals (including Mules) were harmed during this Meetup.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lH-l-urZn4mzFy1Rw1PfCw.jpeg" /><figcaption>Mariana and I with Max the Mule</figcaption></figure><p>A heat and drought advisory had recently been issued and Houston was HOT. But, on the day of the first hybrid MuleSoft Meetup at the Slalom office, it started to pour.</p><p>The city center has a special place in my heart, it’s where I go to wind down. I take a mental health day once every 3 or 4 months and go down to the city center, find a nice cafe, and settle down with a book. I pick the middle of the week because weekends are busy, both in the city center and at home with two kids. I had not been to the city center in a while as a result of both the pandemic and, well, life!</p><p>On this particular day, I was done with my work, dressed and ready, the kids were home with my husband. Did I already mention it started to pour? I was not happy about it; I dislike driving in the rain (psst.. I hate it, but my son does not like it when I use the word hate), and I knew it was going to be a long drive. Houston traffic gets backed up when it rains.</p><p>An hour and a half later, I finally arrived at the city center parking garage, next to the Slalom office. This was my first time at the office, and I was already late. So, instead of walking around trying to figure out where the entrance was, I called Jim Andrews. He picked up my call and greeted me with a cheerful “hello!”. If you have met Jim, you know what I mean. He gave me quick directions to the office. When I came in, Jim, Raj, and Marco were heating pizza.</p><p>I turned around, and there was my dear friend, Mariana. She had come from Mexico with her Husband, Marco, to meet us. This was the first time I was meeting Mariana Lemus, MuleSoft Community Manager, in-person. We have frequent conversations on Slack, Twitter, and LinkedIn, but in person is always special. After a hello and a quick hug, she handed me a t-shirt and asked me to wear it to have some pictures taken.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*baiY4mB6X05Fj5CUv6OdFA.jpeg" /><figcaption>Houston Muleys</figcaption></figure><p>We took a few group pictures and selfies and then settled down for a conversation. Jim was hacking away at the projector to get the set-up going- this was hybrid, remember? We had friends tuning in from all over: AMER, LATAM, and India. This was Meetup number 18 for Houston and we wanted it to be perfect!</p><p>I could not help but reminisce on the first <a href="https://meetups.mulesoft.com/houston/">Houston MuleSoft Meetup</a>. It was different and I was just starting, with only a few projects under my belt. I was a MuleSoft newbie with a desire to make a difference and I gave my first MuleSoft speech on API-led connectivity and how it helps businesses unlock data for reuse. Today, I was here to talk about my journey as a <a href="http://mulesoft.com/mentors">MuleSoft Mentor</a>. It has been an interesting journey, <strong><em>Tempora mutantur, nos et mutamur.</em></strong> Since my first Houston MuleSoft Meetup, I have been nominated as a MuleSoft Mentor, write blogs, and frequently speak at MuleSoft Meetups. On the personal front, I have two kids who make this all worth it and a husband who finds it amusing that I am a public speaker but can’t respond to his pointed question with a straight sentence. I was nudged back to the present by Marco who asked if I wanted a slice of pizza. I said yes!</p><p>Marco and I, along with other attendees, made a beeline to the conference room. I have always found professionalism to be a very powerful quality, and I was excited to present my speech to the Meetup community on the topic “Who are you as a professional?”. To me, it’s a means to inspire, impress, and allows you to do what you are passionate about in the best way possible. The Meetup began with Jim opening with introductions and a talk about Houston Meetups, and the agenda plan. After that, it was Mariana who took the stage, speaking about her recent trips, the MuleSoft mentorship program, and details about the growth of the MuleSoft Community. It was both amazing and humbling. The MuleSoft Community (myself included) is constantly looking for ways to give back because when each of us started, the community was there to help orient, inspire, and encourage us. The MuleSoft mentorship program is tailored to help individuals find the perfect opportunity for them, based on areas of interest and MuleSoft expertise. The program connects like-minded individuals who can benefit from mentorship. You can find more details <a href="https://developer.mulesoft.com/community/mentors">here</a>.</p><p>Next, it was my turn to speak. I believe that if you do not plan your journey right you may not like where you end up. I do not know if I will feel that way in a decade, <strong><em>Tempora mutantur, nos et mutamur </em></strong><em>remember? </em>But, what I do know is that right now I am happy with where my journey has brought me. It has brought me to you - the API and integrations community. It was a very quiet drive for me back home. It had stopped raining and the roads were clear. I felt refreshed, I had enjoyed great company. I learned something new about myself that day- I would like to continue to grow, with all of you, to inspire, lead and educate. I hope that, whenever I look back, I will always feel grateful for the experiences I have had.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=339818023a1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/another-integration-blog/pizza-wine-and-great-company-339818023a1">Pizza, Wine, and Great Company!</a> was originally published in <a href="https://medium.com/another-integration-blog">Another Integration Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>