<?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 Adilakshmi Vadavalli on Medium]]></title>
        <description><![CDATA[Stories by Adilakshmi Vadavalli on Medium]]></description>
        <link>https://medium.com/@adilakshmi.vadavalli?source=rss-67154c86264e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*e45lSyQ6-BLSbAda.jpg</url>
            <title>Stories by Adilakshmi Vadavalli on Medium</title>
            <link>https://medium.com/@adilakshmi.vadavalli?source=rss-67154c86264e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:47:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@adilakshmi.vadavalli/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[Don’t Just Build AI. Move It.]]></title>
            <link>https://medium.com/@adilakshmi.vadavalli/dont-just-build-ai-move-it-a445612a1eb7?source=rss-67154c86264e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a445612a1eb7</guid>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[speed]]></category>
            <category><![CDATA[agentic-ai-application]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[regulatory-compliance]]></category>
            <dc:creator><![CDATA[Adilakshmi Vadavalli]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 14:58:53 GMT</pubDate>
            <atom:updated>2026-11-11T14:58:53.530Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Elephant, the Rider, the Path and the Governance Gates That Decide Whether Agentic AI Reaches Production</h3><p><em>A practical reflection on enterprise Agentic AI delivery, speed, strategic prioritisation, governance and user adoption.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hfq42PGaULHe2VlIn2F6lg.png" /></figure><h3>1. The Elephant Is Ready. But Is the Organisation Ready to Move?</h3><p>Recently, I picked up a copy of <em>Switch: How to Change Things When Change Is Hard</em> by Chip Heath and Dan Heath.</p><p>The book introduces three interesting characters:</p><ul><li><strong>The Elephant:</strong> Our emotional side. It provides motivation and momentum.</li><li><strong>The Rider:</strong> Our rational side. It provides direction and planning.</li><li><strong>The Path:</strong> The environment that makes the journey easier or harder.</li></ul><p>While reading the book, I started connecting these concepts to something I was experiencing in an enterprise Agentic AI project.</p><p>Imagine a team committing to deliver an enterprise AI application with 12 features in three months. The architecture is ready. Developers are building agents, integrating tools, implementing RAG, configuring guardrails and testing workflows.</p><p>Three months become five.</p><p>Five months become seven.</p><p>The application finally reaches QA, only to encounter security reviews, access provisioning, architecture approvals, G1 and G3 governance gates, and production readiness requirements.</p><p>Meanwhile, something interesting happens.</p><p><strong>The end users have started building their own equivalent chatbot!</strong></p><p>Not necessarily because their solution is better, but because they need something that works today.</p><p>That made me realise something important:</p><p><strong>In the AI race, a technically successful application can still lose the adoption race.</strong></p><p>The competition is not always another technology company. Sometimes, it is the end user solving the same problem independently.</p><h3>2. Lessons Learnt: The Challenges That Changed My Perspective</h3><p>Working through this journey made me realise that the biggest challenges in enterprise Agentic AI are not always related to models, frameworks or technical architecture.</p><p>Some of the toughest challenges emerge when we try to move a working application into the hands of real users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C1XCVaBq5I0_J8MZbKs0KQ.jpeg" /></figure><p>Here are four important lessons.</p><p><strong>Lesson 1: Time-to-market can make or break an AI application.</strong></p><p>We initially plan for three months of development. It takes five months to complete development, followed by another three months navigating the journey to production.</p><p>By then, the original business opportunity may have changed.</p><p>The end users may have already explored alternative AI tools or built their own solutions.</p><p>An application that arrives too late may no longer solve the problem it was originally intended to address.</p><p><strong>Lesson 2: User adoption cannot be assumed.</strong></p><p>End users are becoming increasingly aware of AI capabilities.</p><p>They question:</p><ul><li>Why should I use this application?</li><li>Is it better than the chatbot I already use?</li><li>Can it complete my task end to end?</li><li>Can I trust its answers?</li><li>How much time will it actually save?</li></ul><p>A successful demo does not automatically translate into successful adoption.</p><p>We need to demonstrate business value, reliability and trust.</p><p><strong>Lesson 3: Governance and approvals can become critical dependencies.</strong></p><p>Access provisioning, G1 and G3 approvals, security assessments, compliance reviews, infrastructure readiness and production authorisation can introduce significant lead time.</p><p>Development may be progressing, but the path to production remains blocked.</p><p>The challenge is not that governance exists. Governance is necessary.</p><p>The challenge is that we often start planning for governance too late.</p><p><strong>Lesson 4: Completing features is not the same as delivering business value.</strong></p><p>Imagine completing all 12 features in development but not getting the application into production.</p><p>What have we actually achieved?</p><p>A feature sitting in a development environment does not create the same value as a feature being used by real employees to solve a business problem.</p><p>This brings us to a fundamental shift in thinking:</p><p><strong>We should measure success by the value delivered and adopted, not merely by the number of features developed.</strong></p><h3>3. The Wake-up Call: Speed Is No Longer Just a Delivery Metric. It Is a Survival Strategy.</h3><p>Agentic AI is evolving rapidly.</p><p>Models, frameworks, orchestration patterns and user expectations are changing continuously.</p><p>The window of opportunity can be surprisingly short.</p><p>Traditional application delivery approaches often focus on completing the entire scope before a major release.</p><p>But in the Agentic AI world, should we really wait for all 12 features before releasing anything?</p><p>What if we identify the smallest four features that solve a meaningful business problem?</p><p>Take them through development, QA, security validation and the required governance approvals. Release them. Collect real user feedback.</p><p>Then deliver two more features.</p><p>Continue incrementally.</p><p><strong>Build small. Release early. Learn continuously. Improve relentlessly.</strong></p><h3>4. The Strategy: Stop Building 12 Features. Start Delivering the First Four.</h3><p>This is where I see a direct connection with the <em>Switch</em> principle of shrinking the change.</p><p>Suppose we have 12 features in our roadmap.</p><p>Instead of attempting to deliver everything together, prioritise the minimum set of capabilities that provides genuine business value.</p><p>The actual number of features in each release should depend on dependencies, business value, risk and production readiness.</p><p>Why does this approach work?</p><ul><li>Earlier business value.</li><li>Faster feedback from actual users.</li><li>Smaller testing and deployment scope.</li><li>Earlier identification of security and governance gaps.</li><li>Better stakeholder confidence.</li><li>Reduced risk of building features nobody eventually uses.</li></ul><p>One important distinction:</p><p><strong>Small releases do not mean small governance.</strong></p><p>Every release must satisfy its applicable security, privacy, quality and operational requirements.</p><p>We should reduce unnecessary scope, not necessary controls.</p><h3>5. The Elephant, the Rider and the Path: Applying the Switch Model to Agentic AI</h3><p>The book introduces three powerful ways to make change happen. Let’s apply them to enterprise AI delivery.</p><p><strong>The Rider — Give the project a clear direction.</strong></p><p>The Rider represents planning and rational decision-making.</p><p>In an Agentic AI project, the Rider should answer:</p><ul><li>What is the primary business problem?</li><li>What is the minimum viable product?</li><li>Which four features deliver maximum value?</li><li>What are the technical and organisational dependencies?</li><li>Who owns each approval?</li><li>What is the path from development to production?</li></ul><p>A roadmap containing 12 features is not necessarily a strategy.</p><p>A strategy also tells us what NOT to build right now.</p><p><strong>The Elephant — Create momentum.</strong></p><p>The Elephant represents motivation and emotional energy.</p><p>Development teams lose momentum when they spend months building without seeing their work reach actual users.</p><p>Stakeholders lose confidence when milestones keep moving.</p><p>Users lose interest when they keep waiting.</p><p>Celebrate meaningful production releases, not just development completion.</p><p>A feature being actively used by 50 employees can be more valuable than 12 features sitting in development.</p><p>Small wins create confidence. Confidence creates momentum.</p><p><strong>The Path — Remove organisational obstacles.</strong></p><p>This is perhaps the most underestimated part of enterprise AI delivery.</p><ul><li>Are access requests initiated early?</li><li>Are security and architecture teams involved from day one?</li><li>Are G1 and G3 requirements understood?</li><li>Is the infrastructure provisioned?</li><li>Are QA environments available?</li><li>Are business owners ready to validate?</li><li>Is there a named production owner?</li></ul><p>If these activities begin only after development is complete, we have already created a bottleneck.</p><p><strong>The production path must be designed alongside the application architecture.</strong></p><h3>6. G1 and G3: Connecting Enterprise Governance to India’s AI Governance Ecosystem</h3><p>This is where enterprise Agentic AI delivery becomes particularly interesting.</p><p>G1 and G3 are not universal regulatory gate names. Their exact definitions depend on an organisation’s internal project governance framework.</p><p>For this discussion, let’s use an illustrative interpretation:</p><ul><li><strong>G1 — Project initiation and risk readiness.</strong></li><li><strong>G3 — Release and production readiness.</strong></li></ul><p>We can map these enterprise stage gates to recognised governance and security frameworks such as:</p><ol><li>India’s AI Governance Guidelines (MeitY).</li><li>Digital Personal Data Protection Act, 2023 and DPDP Rules, 2025.</li><li>NIST AI Risk Management Framework.</li><li>OWASP Top 10 for Agentic Applications.</li><li>CERT-In cybersecurity directions, wherever applicable.</li></ol><p>The objective is not to create additional paperwork. It is to translate governance expectations into engineering activities and measurable evidence.</p><h4>G1 — Project initiation and risk readiness</h4><p>Think of G1 as answering:</p><p><strong>Should we build this, and under what conditions?</strong></p><p>Before investing months in development, establish whether the proposed solution is justified, feasible, accountable and sufficiently understood.</p><p><strong>A. India AI Governance Guidelines — MeitY</strong></p><p>India’s AI Governance Guidelines introduce seven guiding principles:</p><ul><li>Trust is the foundation.</li><li>People first.</li><li>Innovation over restraint.</li><li>Fairness and equity.</li><li>Accountability.</li><li>Understandable by design.</li><li>Safety, resilience and sustainability.</li></ul><p>At G1, translate these principles into practical questions.</p><ul><li>Who benefits from this application?</li><li>Who could be harmed?</li><li>Who is accountable for the agent’s decisions and actions?</li><li>Is human oversight necessary?</li><li>Can users understand the system’s limitations?</li><li>What happens when the agent makes a mistake?</li></ul><p><strong>B. DPDP Act, 2023 and DPDP Rules, 2025</strong></p><p>If the Agentic AI application processes personal digital data, privacy must be considered before development.</p><p>At G1, identify:</p><ul><li>What personal data will be accessed?</li><li>Why is it required?</li><li>What is the applicable lawful basis for processing?</li><li>Are purpose limitation and data minimisation addressed?</li><li>Who can access the information?</li><li>How will data be protected and retained?</li><li>How will applicable data principal rights be supported?</li></ul><p>For example, imagine an HR assistant that accesses employee records.</p><p>Do we really need to expose the entire employee database to the agent?</p><p>Or can we restrict its access to only the information required for a specific task?</p><p>That is privacy by design.</p><p>The DPDP Rules were notified in November 2025 with phased commencement. Organisations must map their obligations to the applicable commencement dates.</p><p><strong>C. NIST AI Risk Management Framework</strong></p><p>NIST AI RMF provides four functions:</p><p>FunctionApplication at G1GovernDefine ownership, policies and accountabilityMapIdentify business context, stakeholders and risksMeasureEstablish evaluation criteria and risk metricsManageDefine risk treatment and mitigation plans</p><p>For example, before developing an autonomous procurement agent, identify the potential financial, operational, privacy and security risks.</p><p>NIST provides a lifecycle-oriented approach rather than treating AI risk management as a one-time approval.</p><p><strong>D. OWASP Top 10 for Agentic Applications</strong></p><p>Agentic AI introduces risks beyond traditional chatbot development.</p><p>Consider:</p><ul><li>Agent goal hijacking.</li><li>Tool misuse.</li><li>Identity and privilege abuse.</li><li>Agentic supply-chain vulnerabilities.</li><li>Unexpected code execution.</li><li>Memory and context poisoning.</li><li>Insecure inter-agent communication.</li><li>Cascading failures.</li><li>Human-agent trust exploitation.</li><li>Rogue agent behaviour.</li></ul><p>At G1, identify which risks are relevant to the architecture and decide what security controls must be built.</p><p>For example, if an agent can raise purchase orders, should it have unrestricted access to the procurement system?</p><p>Absolutely not.</p><p>Define permitted tools, least-privilege access, approval boundaries and audit requirements before implementation.</p><p><strong>G1 expected outcome:</strong></p><p>An approved business problem, defined scope, initial risk assessment, ownership, architecture direction, data classification and documented governance requirements.</p><h4>G3 — Release and production readiness</h4><p>Now imagine the application has been developed and has passed functional testing.</p><p>Can we immediately move it to production?</p><p>Not necessarily.</p><p>G3 should establish whether the application is safe, secure, reliable, compliant and operationally ready for its intended use.</p><p>Think of G3 as answering:</p><p><strong>Can we responsibly release this application to real users?</strong></p><p><strong>A. India AI Governance Guidelines</strong></p><p>Validate:</p><ul><li>Is human oversight implemented where required?</li><li>Is accountability clearly assigned?</li><li>Are users informed about AI limitations?</li><li>Are fairness and reliability evaluated?</li><li>Is there a process to handle complaints and incidents?</li></ul><p><strong>B. DPDP Act and Rules</strong></p><p>Validate applicable privacy obligations:</p><ul><li>Are personal data access controls implemented?</li><li>Are appropriate safeguards in place?</li><li>Are purpose and retention requirements addressed?</li><li>Can applicable data principal rights be supported?</li><li>Is there a personal data breach response process?</li><li>Are responsibilities assigned?</li></ul><p>Privacy approval is not simply a document-signing exercise. It must be reflected in actual application behaviour.</p><p><strong>C. NIST AI RMF</strong></p><p>FunctionG3 production evidenceGovernApproved ownership, policies and accountabilityMapDocumented use cases, limitations and residual risksMeasureEvaluation results, reliability and security testingManageRisk mitigation, monitoring and incident response</p><p>For Agentic AI, evaluation should go beyond answer accuracy.</p><p>We should test tool selection, execution permissions, multi-agent behaviour, failure recovery, human approval mechanisms and unintended actions.</p><p><strong>D. OWASP Agentic AI security</strong></p><p>Validate the actual implementation:</p><ul><li>Can prompt injection manipulate an agent?</li><li>Can an agent invoke an unauthorised tool?</li><li>Can sensitive data leak through responses?</li><li>Can one agent manipulate another agent’s instructions?</li><li>Are tool calls authenticated and authorised?</li><li>Are actions logged and traceable?</li><li>Can high-impact actions require human approval?</li></ul><p>A successful functional test does not establish that an agent is safe to deploy.</p><p><strong>E. CERT-In cybersecurity directions</strong></p><p>For organisations within scope, production readiness should also consider applicable CERT-In directions.</p><p>The 2022 directions include requirements relating to:</p><ul><li>Reporting specified cyber incidents within six hours of noticing them or being informed of them.</li><li>Enabling ICT system logs and securely maintaining them for a rolling 180-day period within Indian jurisdiction.</li></ul><p>This makes logging, incident response, time synchronisation and ownership important operational considerations.</p><p><strong>G3 expected outcome:</strong></p><p>Evidence of completed testing, resolved mandatory risks, appropriate approvals, access controls, monitoring, incident response, rollback readiness and an accountable production owner.</p><p>The important distinction is this:</p><p><strong>G1 establishes whether we should proceed and what controls are needed. G3 establishes whether we are ready to release responsibly.</strong></p><p>The exact approval evidence and signatories must always follow the client’s internal gate framework.</p><h3>7. The Real Strategy: Run Development and Governance in Parallel</h3><p>Here is where I would change the conventional delivery approach.</p><p>Instead of:</p><p>Development → QA → Security → Governance → Access → Production</p><p>Think about parallel workstreams.</p><p><strong>Workstream 1: Product engineering</strong></p><ul><li>Prioritise the first four features.</li><li>Build and test incrementally.</li><li>Implement guardrails and evaluations.</li><li>Prepare subsequent releases.</li></ul><p><strong>Workstream 2: AI governance and security</strong></p><ul><li>Identify G1 and G3 requirements upfront.</li><li>Map DPDP obligations.</li><li>Apply NIST AI RMF.</li><li>Use OWASP guidance to identify and mitigate agentic risks.</li><li>Maintain evidence throughout development.</li></ul><p><strong>Workstream 3: Infrastructure and access</strong></p><ul><li>Initiate access requests early.</li><li>Provision environments.</li><li>Validate identity and network connectivity.</li><li>Prepare production infrastructure.</li></ul><p><strong>Workstream 4: User adoption</strong></p><ul><li>Identify a small group of real users.</li><li>Demonstrate the application early.</li><li>Conduct user acceptance testing.</li><li>Collect feedback and measure adoption.</li></ul><p>This way, we are not waiting for development to finish before thinking about production.</p><p>We are preparing the production journey while building the product.</p><p>Of course, parallel preparation does not mean bypassing mandatory approval dependencies.</p><h3>8. A Question Every Agentic AI Project Leader Should Ask</h3><p>Before committing to a roadmap, I would ask:</p><ol><li>What is the smallest meaningful business problem we can solve?</li><li>Do we really need all 12 features in the first release?</li><li>How quickly can we move the first usable capability to production?</li><li>Have we identified G1, G3 and other approval dependencies?</li><li>Who is accountable if an agent makes an incorrect decision or performs an unintended action?</li><li>Are privacy, security, access controls and guardrails implemented?</li><li>Are end users involved throughout development?</li><li>How quickly can we convert feedback into the next release?</li></ol><p>And perhaps the most important question:</p><p><strong>Are we building an application because the technology is exciting, or are we solving a problem that someone genuinely needs solved?</strong></p><p>That distinction can determine whether an AI application becomes a business asset or another abandoned experiment.</p><h3>9. Five Key Takeaways</h3><p><strong>1. Shrink the change.</strong></p><p>Don’t wait for 12 features. Deliver the smallest valuable four.</p><p><strong>2. Speed creates survival.</strong></p><p>Delayed delivery can mean lost adoption and lost business relevance.</p><p><strong>3. Governance starts on Day 1.</strong></p><p>Plan G1, G3, DPDP, NIST, OWASP, access and infrastructure alongside development.</p><p><strong>4. Production is not the finish line.</strong></p><p>User adoption, feedback, trust and continuous improvement determine long-term success.</p><p><strong>5. Momentum is everyone’s responsibility.</strong></p><p>Product, engineering, security, governance, infrastructure and business teams must move together.</p><h3>10. Finally, Let’s Get the Elephant Moving!</h3><p>I started reading <em>Switch</em> as a book about organisational and personal change.</p><p>But my recent experience with enterprise Agentic AI made me look at it differently.</p><p>Perhaps the biggest challenge in enterprise AI isn’t just building intelligent agents.</p><p>It is aligning people, processes, technology, governance and business priorities to get those agents into the hands of users.</p><p>We need the Rider to think strategically.</p><p>We need the Elephant to keep moving.</p><p>And we need to shape the Path so that unnecessary obstacles don’t slow us down.</p><p>I would rather see four meaningful features solving a real business problem in production today than twelve beautifully engineered features waiting for their turn.</p><p>Because the AI landscape will not wait for our perfect roadmap.</p><p><strong>Build small. Move fast. Govern intelligently. Learn continuously. And never lose momentum.</strong></p><p>The real measure of an Agentic AI project’s success isn’t how many agents we build.</p><p><strong>It’s how much meaningful value we deliver before the opportunity moves on.</strong></p><h3>References and Further Reading</h3><ol><li>Chip Heath and Dan Heath — <em>Switch: How to Change Things When Change Is Hard</em>.</li><li><a href="https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc2025115685601.pdf">India AI Governance Guidelines — MeitY, Government of India</a>.</li><li><a href="https://www.meity.gov.in/documents/act-and-policies/digital-personal-data-protection-rules-2025-gDOxUjMtQWa">Digital Personal Data Protection Act and Rules — MeitY</a>.</li><li><a href="https://www.nist.gov/itl/ai-risk-management-framework">NIST AI Risk Management Framework</a>.</li><li><a href="https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/">OWASP Top 10 for Agentic Applications 2026</a>.</li><li><a href="https://www.cert-in.org.in/Directions70B.jsp">CERT-In Cybersecurity Directions</a>.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a445612a1eb7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Silent Errors in Serverless: The Failures Nobody Sees]]></title>
            <link>https://medium.com/@adilakshmi.vadavalli/silent-errors-in-serverless-the-failures-nobody-sees-e6293b51b772?source=rss-67154c86264e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6293b51b772</guid>
            <dc:creator><![CDATA[Adilakshmi Vadavalli]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 06:58:12 GMT</pubDate>
            <atom:updated>2026-07-13T06:58:12.136Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Silent Errors in Serverless: The Failures Nobody Sees</strong></p><p><em>How invisible failures in cloud-native architectures silently corrupt your data — and 5 practices to make them loud.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9IjyA23rhb70z0yVIrAshQ.jpeg" /></figure><p>Your Lambda function returns 200. Your Step Functions execution shows ‘Succeeded’. Your CloudWatch dashboard is all green. But your user is looking at data from three days ago.</p><p>Welcome to silent errors — the most insidious class of bugs in serverless architectures.</p><h3>What is a Silent Error?</h3><p>A silent error is a failure that doesn’t crash your system, doesn’t trigger an alarm, and doesn’t return an error to the caller. The operation fails, but execution continues as if nothing happened. The database doesn’t get updated. The state doesn’t change. But everything downstream assumes it did.</p><p>In traditional monolithic systems, these failures usually manifest as visible crashes. But in serverless — where every operation is wrapped in retry logic, error handlers, and fail-open patterns — they become invisible.</p><h3>The Anatomy of a Silent Failure</h3><p>Every silent error I’ve encountered in production follows the same five-step lifecycle:</p><p>· A developer adds a try/except around a database write as a ‘safety net’.</p><p>· An unexpected failure occurs (wrong type, reserved word, missing permission).</p><p>· The generic exception handler catches it and logs a warning.</p><p>· The function returns success to the caller.</p><p>· The user sees stale data. Nobody knows for days.</p><p>The tragedy is that the developer had good intentions. They didn’t want a failed database update to crash the entire request. But by absorbing the failure, they created something worse: a system that lies about its own state.</p><h3>Three Silent Failures commonly Found in Production</h3><h3>1. The Reserved Word That Ate My Updates</h3><p>Most managed databases have hundreds of reserved keywords. DynamoDB alone has 573. When you use one as a field name in an update expression — ‘status’, ‘name’, ‘sources’, ‘count’ — the operation fails. But if it’s inside a try/except, the failure vanishes.</p><p>In our case, a field called ‘sources’ was a DynamoDB reserved word. Every update to that field silently failed. The API returned 200. The user saw the old data. It took three days to find because the error was logged as a WARNING — and nobody had a CloudWatch alarm on warnings.</p><h3>2. The Float That Wasn’t a Decimal</h3><p>When you read JSON (from an API, from S3, from anywhere), numbers come back as floats. When you write to DynamoDB using the boto3 resource layer, it rejects floats — it wants Decimal objects. This type mismatch throws a TypeError, which gets caught by your generic exception handler, and your write never happens.</p><p>The insidious part: this only fails on records that HAVE decimal numbers. If your test data uses whole numbers (integers), the tests pass. The bug only surfaces in production with real data that contains relevance scores like 0.85 or confidence values like 0.7.</p><h3>3. The Exception Handler That Hid Everything</h3><p>This is the enabler of all silent errors. The pattern looks innocent:</p><p>try:<br> db.update(key, data)<br> except Exception as e:<br> logger.warning(f’update failed: {e}’)</p><p>The developer’s intent: ‘if the update fails, don’t crash — log it and move on.’ The reality: every TypeError, every reserved word error, every permission denial is caught, whispered into a log file nobody reads, and forgotten. The system lies to every downstream consumer.</p><h3>5 Practices That Eliminate Silent Failures</h3><h3>1. Always Escape Field Names</h3><p>Never use raw field names in database expressions. Period. Even if you believe the name isn’t reserved today, future SDK updates may add new reserved words. Treat parameterized field references as a universal default, not an exception. The extra 2 lines of code save you 2 days of debugging.</p><h3>2. Read After Write for Critical Operations</h3><p>For state transitions that users depend on (status changes, permission grants, financial updates), add a verification read immediately after the write. If what you just wrote doesn’t match what you read back, raise an error loudly. This catches silent failures at the moment they happen — not days later when a user reports wrong data.</p><h3>3. Use Specific Exception Types</h3><p>Replace every ‘except Exception’ with specific types: ClientError for database issues, TypeError for serialization problems, ValidationError for schema violations. Let unexpected exceptions propagate and crash. A crash you see immediately is infinitely better than a silent corruption you discover next week.</p><h3>4. Log Observable State Transitions</h3><p>Don’t log ‘update successful.’ Log the measurable state change: ‘sources: 14 → 11’ or ‘status: PENDING → APPROVED.’ If your logs show ’14 → 14&#39; (no change), you immediately know the write failed silently. Numbers don’t lie. Generic success messages do.</p><h3>5. Test Real Writes Locally Before Deploying</h3><p>Before deploying any code that writes to a managed database, run the exact operation locally against the real table with real data. Not mocked. Not simulated. The actual SDK call with actual data shapes that exist in production. Most silent failures — reserved words, type mismatches, missing fields — surface instantly in a 5-second local test but hide for weeks behind generic error handlers in production.</p><h3>The Core Principle</h3><p>In serverless architectures, the most dangerous bug is not the one that crashes your system — it’s the one that doesn’t.</p><p>Silent errors erode data integrity invisibly. They accumulate. They compound. By the time someone notices, the damage is spread across hundreds of records, multiple downstream services, and weeks of user interactions built on corrupted state.</p><p><strong>Build systems that fail loudly or succeed verifiably. There is no safe middle ground.</strong></p><p>· Reserved words in database expressions → always escape field names</p><p>· Type mismatches (float vs Decimal) → convert explicitly before writes</p><p>· Broad exception handlers → use specific types, let unknowns crash</p><p>· No verification → read-after-write for critical operations</p><p>· No local testing → run real writes against real tables before deploy</p><p><em>The cost of a loud failure: a 5-minute fix.<br> The cost of a silent failure: days of debugging, lost user trust, and data you can never fully recover.</em></p><p>#Serverless #AWS #DynamoDB #Lambda #CloudArchitecture #DevOps #SilentErrors #BestPractices</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e6293b51b772" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Demo to Deployment]]></title>
            <link>https://medium.com/@adilakshmi.vadavalli/from-demo-to-deployment-64e44a9d1641?source=rss-67154c86264e------2</link>
            <guid isPermaLink="false">https://medium.com/p/64e44a9d1641</guid>
            <dc:creator><![CDATA[Adilakshmi Vadavalli]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 04:45:21 GMT</pubDate>
            <atom:updated>2026-07-06T06:13:15.758Z</atom:updated>
            <content:encoded><![CDATA[<p>The Reality Gap Enterprises Need to Understand About Agentic AI</p><p><em>Why a working prototype and a production system are two entirely different engineering problems — told through seven diagrams</em></p><h3>Introduction: Why AI POCs Look Easy but Production AI Is Hard</h3><p>It always starts the same way.</p><p>A small team gets excited about an creating a chatbot interface for a client. They connect an LLM to a set of documents, wire up a chatbot interface, add a couple of tools, and within two weeks they’re demoing something that genuinely feels like magic. The room is impressed. Someone in leadership asks the inevitable question:</p><p><em>“This looks done. Why can’t we ship it next week?”</em></p><p>It’s a fair question — and it’s also the moment where most AI initiatives quietly start to struggle.</p><p>Because a POC only has to answer one question: can this idea technically work? Production has to answer a much harder one: can this operate safely, every day, for thousands of users, without falling over, leaking data, blowing the budget, or hallucinating its way into a compliance incident?</p><p>A demo proves possibility. Production proves engineering maturity. Everything below is the story of that gap — chapter by chapter, diagram by diagram.</p><h3>POC Goal vs Production Goal</h3><p>A Proof of Concept mainly answers: “Can this idea technically work?” The focus is usually to demonstrate business value, validate model capability, show the user experience, and prove feasibility — with limited users, limited documents, controlled inputs, and manually selected examples. Success is measured by possibility.</p><p>Production systems answer a different question: “Can this operate safely every day for thousands of users?” The focus shifts to reliability, security, scalability, compliance, monitoring, cost optimization, and maintainability. Success is measured by operational excellence.</p><h3>1. Enterprise Infrastructure Constraints: AI Meets Reality</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_6TcNVBj0cj-g82vhz9WSA.png" /></figure><p>In the POC, the architecture is almost embarrassingly simple: User → Application → LLM → Response. A public API key, an open network connection, a developer’s laptop. Fast, frictionless, and completely unrepresentative of where the system will actually live.</p><p>The moment that same idea walks into an enterprise, it meets a very different world. There’s no more “just call the API.” Every request now has to pass through identity management, an API gateway, a security layer, an agent orchestrator, and only then does it reach an approved, private model endpoint — before flowing back through monitoring and governance.</p><p>This is where teams discover the real challenge isn’t accessing AI. It’s safely integrating it — inside VPCs, behind private endpoints, under model-approval processes, respecting data residency rules that never came up during the demo. Long approval cycles, restricted internet access, and mandatory audit logging aren’t red tape for its own sake; they’re what “production-grade” actually means before a single token gets processed.</p><h3>2. Guardrails and Agentic Workflows: Controlling Agent Behaviour at Scale</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y4ZnhvQyHfW-ZO_O7NGw8w.png" /></figure><p>There’s a second, quieter shift happening at the same time: the AI itself gets more complex.</p><p>A POC agent looks like this: User Question → LLM → Tool Call → Answer. One hop, one tool, one shot. Production agents don’t get that luxury. Real agentic workflows involve a planning layer, a reasoning engine, tool selection, security validation, execution, observation, and reflection — before a response is even generated. Every one of those extra steps is a place where cost, latency, and failure can quietly creep in.</p><p>And because the agent is now doing things, not just saying things, it needs guardrails at every layer it operates in — system-level rules that define its identity and boundaries, application-level checks that enforce business logic, environment-specific tightening as it moves from development to staging to production, and continuous monitoring that watches for anomalies in real time. Every action ultimately resolves to one of four outcomes: allow it, sanitize it, block it, or escalate it to a human. That decision tree, running silently behind every agent step, is the difference between an agent you can trust with real systems and one that’s still just a clever demo.</p><h3>3. Token Economy: Managing AI Cost Engineering</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U2_JyODeZI1FG_YuiIPy3w.png" /></figure><p>In a POC, nobody worries about tokens. “Just send everything to the LLM” is a perfectly reasonable strategy when you’re running ten test queries a day.</p><p>At scale, that same habit becomes a line item that finance notices. Every input token, every reasoning token, every retrieval and reranking pass, every tool call, every retry after a failure — it all adds up, and it adds up fast when thousands of users are hitting the system concurrently.</p><p>Production teams end up building an entire token optimization pipeline: compressing prompts, retrieving only what’s relevant, deduplicating and caching aggressively, routing requests to the right-sized model instead of the biggest one by default, and continuously monitoring usage and cost. The mindset shift is simple to say and hard to build: send only what is necessary — and treat every token as an engineering decision, not an afterthought.</p><h3>4. Context Engineering and RAG Architecture Design</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BrABY7I94djWCmM69eymvw.png" /></figure><p>Early on, most teams lean entirely on prompt engineering — pack enough instructions and examples into the prompt and hope for the best. It works, right up until the prompt becomes enormous, the rules start contradicting each other, and responses become inconsistent in ways nobody can quite explain.</p><p>The real fix isn’t a better prompt. It’s context engineering — deciding what information the model needs, when it needs it, how much is enough, and which source can actually be trusted.</p><p>This is most visible in Retrieval-Augmented Generation. A demo RAG pipeline is refreshingly simple: chunk the documents, embed them, run a similarity search, hand the results to the LLM. It works well enough to impress a room.</p><p>Enterprise RAG is a different animal entirely — document intelligence, semantic chunking, metadata enrichment, hybrid retrieval, reranking, and careful context assembly, all wrapped in governance for access control, document permissions, staleness detection, and citation accuracy. The goal was never more context. It was always the right context — grounded, permissioned, and current.</p><h3>5. Non-Deterministic AI Systems: Testing the Untestable</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7jbnXbmc_nmTRNDK_dlrxQ.png" /></figure><p>Traditional software gives you a comforting guarantee: same input, same output. Test it once, and you’ve tested it forever.</p><p>AI systems break that promise on purpose. The same query can produce several different, equally valid responses — and that single fact quietly undermines how QA has worked for decades. How do you reproduce a failure that might not happen again? How do you measure “correctness” when there isn’t one correct answer?</p><p>The answer is a genuine evaluation layer sitting between the model and the decision to trust its output — automated metrics, LLM-as-a-judge scoring, rule-based checks, consistency and groundedness checks, safety and policy checks. From there, the system decides: accept the response, retry it, re-rank it, refine it, escalate it to a human, or fall back to a safer default. Every one of those decisions gets fed back into a continuous loop of collecting feedback, analyzing trends, and retuning prompts, context, and parameters. Non-determinism isn’t a bug to eliminate — it’s a property to be managed, measured, and continuously improved against.</p><h3>6. Latency and Caching: Making AI Production Ready</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-ihvbGtEbD6pNKI7v4mEOg.png" /></figure><p>A POC user will forgive a 15-second wait — they’re impressed the thing works at all. A production user will not. They expect fast, predictable responses, and agentic workflows work directly against that expectation: every reasoning step, retrieval call, tool execution, and validation pass adds latency the demo never had to account for.</p><p>This is where caching stops being an optimization and becomes core architecture. Requests flow through intent detection and a cache lookup before anything expensive happens — a cache hit serves instantly; a miss falls through to context retrieval, tool execution, and inference. Behind that single lookup sits a whole stack of caching layers: edge/CDN caching, semantic caching, vector caching, response caching, prompt template caching, and tool result caching, each catching a different kind of repeated work.</p><p>Layer on parallelization, streaming, smart model routing, connection pooling, warm starts, and batching, and the system starts to feel instant — not because the AI got smarter, but because the engineering around it got serious.</p><h3>7. Safety, Security and Compliance — and the Discipline That Never Stops</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VwcAplTubU6FfotJWvOppw.png" /></figure><p>Everything above eventually has to live inside a lifecycle, not a one-time launch. Plan and design with a threat model in mind. Handle data and context with minimization and consent. Build and test for bias and adversarial inputs. Deploy behind proper access controls and hardened infrastructure. Operate with real monitoring and incident response. Improve continuously through red-teaming and policy updates — and repeat.</p><p>This is also where the last piece of the production story shows up: observability. Traditional monitoring watches CPU, memory, and exceptions. AI monitoring has to watch prompt behavior, model drift, hallucination rates, token usage, retrieval quality, safety violations, and user satisfaction — the discipline generally known as LLMOps, and it’s non-negotiable once real users and real compliance frameworks (GDPR, HIPAA, PCI DSS, ISO 27001, SOC 2, the EU AI Act) are in the picture.</p><p>None of this is about slowing AI down. It’s what “Safety First, Human Oversight, Fairness, Privacy by Design, Transparency, Accountability” actually look like once they stop being principles on a slide and start being architecture.</p><h3>The Final Perspective</h3><p>A POC demonstrates intelligence. Production demonstrates engineering maturity.</p><p>Getting from one to the other isn’t about a better prompt or a bigger model — it’s the sum of infrastructure engineering, guardrails and agentic design, cost engineering, context and retrieval architecture, evaluation science, latency engineering, and security and compliance, all working together, continuously.</p><p>The next generation of AI teams won’t be judged by how fast they can build a demo. They’ll be judged by how reliably they can turn that demo into something thousands of people can trust every single day.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64e44a9d1641" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>