<?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 Pratyosh Desaraju on Medium]]></title>
        <description><![CDATA[Stories by Pratyosh Desaraju on Medium]]></description>
        <link>https://medium.com/@pratyosh.desaraju?source=rss-e0ec3481c855------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Bw1gWIbEfdZuhsr6R2DQFQ.png</url>
            <title>Stories by Pratyosh Desaraju on Medium</title>
            <link>https://medium.com/@pratyosh.desaraju?source=rss-e0ec3481c855------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 22:24:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pratyosh.desaraju/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[Pratyosh Desaraju | Rescuing U.S. Tech Debt: Why Legacy Systems Need AI]]></title>
            <link>https://medium.com/@pratyosh.desaraju/rescuing-u-s-tech-debt-why-legacy-systems-need-ai-2ae36b2ca26c?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/2ae36b2ca26c</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Sat, 15 Feb 2025 12:00:47 GMT</pubDate>
            <atom:updated>2026-10-09T18:47:46.885Z</atom:updated>
            <content:encoded><![CDATA[<p>Over a decade ago, I walked into my first day as an engineer at an insurance company with high expectations. I imagined cutting edge infrastructure, the latest technologies, engineers working on problems that pushed the boundaries of what was possible. Instead, I found myself staring at closets. Entire floors lined with filing cabinets and document storage containing policy records from the 1990s.</p><p>I was genuinely confused.</p><p>A senior engineer laughed and told me the story. They had digitized the documents years ago, scanned everything, built systems to index them. But they could not get rid of the software systems that managed them. The engineers who built those systems had long since retired. Nobody currently working there fully understood how they operated. More importantly, these systems were integrated. Modernizing one would cascade through the entire stack, breaking things that seemed to work fine.</p><p>So they left it all running. Why fix what is not broken, right?</p><p>That thought stuck with me for over a decade. It ignited something I see everywhere in enterprise technology: the tragedy of legacy systems. They are not evil. They are not incompetent design. They are the inevitable result of pragmatic decisions made over time, each one reasonable in its moment, that collectively create a system nobody wants to touch.</p><p>Technical debt is what happens when businesses prioritize speed over architecture. They build systems quickly to meet immediate business needs, promising themselves they will modernize later. Twenty years pass. That “later” never arrives, and suddenly you have a 400 pound gorilla sitting on your IT budget, growing fatter every year with maintenance costs, security patches, and the specialized expertise required to keep it running.</p><p>In the U.S. insurance and retail sectors, legacy systems are not exceptions. They are the foundation upon which entire industries operate. Insurance companies still process claims on systems built in the 1980s. Retailers manage inventory with technology that predates e-commerce by decades. Banks run transactions on mainframes that few people today understand. And the costs? Astronomical. Organizations are spending 70 to 80 percent of their IT budgets just keeping these systems alive, leaving precious little for innovation.</p><p>The problem with legacy systems goes deeper than just being outdated. These systems create data silos where departments cannot speak to each other, where marketing data is siloed from sales data, where inventory data does not communicate with supply chain systems. They are slow, processing transactions in batch overnight rather than in real-time. They are inflexible, requiring months to add a simple new field to a database schema. They are expensive to maintain, requiring specialized expertise that is becoming increasingly rare. They are also security nightmares. A breach in one of these systems could expose millions of customer records, leaving organizations vulnerable to compliance violations.</p><p>But here is where artificial intelligence enters the picture and changes the equation entirely.</p><p>Instead of ripping out your legacy systems which would cost millions, take years, and risk operational disaster AI offers a pragmatic path to awaken them. AI can sit on top of legacy infrastructure and do the heavy lifting that old systems were never designed for. It acts as a translation layer, an intelligence layer, a modernization layer without requiring you to replace the entire foundation.</p><p>In insurance, the impact is transformative. Manual claims processing that currently takes weeks can be automated and completed in days. Machine learning algorithms can assess risk far more accurately than human underwriters, analyzing thousands of data points to determine appropriate premiums. Fraud detection systems can identify anomalies in real-time that human reviewers would miss, saving millions in fraudulent payouts. The system learns continuously, getting smarter with each claim processed.</p><p>Customer service transforms too. Instead of callers waiting on hold for hours to speak with an agent about a claim, AI-powered chatbots handle routine inquiries instantly. Human agents are freed from answering the same questions repeatedly and can focus on complex interactions where their expertise truly matters.</p><p>In retail, demand forecasting moves from educated guesses to AI-powered prediction that factors in weather patterns, social media trends, local events, and competitor activity. Inventory inefficiencies vanish. Stockouts that cost lost sales disappear. Overstocking becomes rare. Personalization becomes possible at scale, with customers seeing recommendations powered by AI analyzing their behavior.</p><p>The beauty of this approach is that you do not need a big bang migration. AI can run alongside your legacy systems in a hybrid model. Your claims processing continues operating as always, while AI gradually takes over tasks, improves efficiency, and proves its value.</p><p>The costs of inaction are substantial. Every year that organizations delay AI integration, they fall further behind competitors already capturing these benefits.</p><p>The U.S. economy runs on these old systems. Insurance companies are managing trillions in policies on infrastructure that was state of the art in 1995. Retailers are handling billions in transactions on systems never designed for real-time personalization. The gap between what these systems can do and what customers expect widens every day.</p><p>AI does not solve technical debt. It manages it. It transforms your legacy systems from anchors into assets that continue generating value while you work toward modernization.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2ae36b2ca26c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Networking for Introverts]]></title>
            <link>https://medium.com/@pratyosh.desaraju/networking-for-introverts-b400ac696c19?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/b400ac696c19</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Sat, 24 Aug 2024 12:00:08 GMT</pubDate>
            <atom:updated>2026-10-09T18:47:27.098Z</atom:updated>
            <content:encoded><![CDATA[<h4>Finding Your Voice Amid the Crowd</h4><p>As an engineer, we have spent more time talking to compilers than humans. My most successful conversations involved debugging sessions with code that at least had the decency to tell me exactly what was wrong. People, on the other hand, are unpredictable.</p><p>I still remember my first big tech conference a few years ago. There I was, standing in a hall filled with 500 engineers who all seemed to know each other. Everyone else was exchanging business cards like they were trading Pokémon cards. Meanwhile, I was calculating the exact distance to the nearest exit.</p><p>Here is the thing no one tells you: most engineers are introverts. We are wired to solve problems, not small talk about the weather.</p><p>And that is okay.</p><p><strong>The Networking Myth<br></strong>We have been sold a story that networking means being loud, charismatic, and somehow able to remember 50 names after one coffee break. That is not networking. That is a superhuman memory test.</p><p>Real networking is about connection. It is about finding one person who cares about the same obscure technical problem you do. It is about depth, not breadth.</p><p><strong>A couple of things I have learned in these networking sessions<br>1. Quality Over Quantity</strong><br> You do not need 100 business cards that lead nowhere because you never really spoke. You need three meaningful conversations. Find the session on the topic you care about most. Show up early. Talk to the person sitting next to you. Ask them what brought them there. Listen.</p><p><strong>2. Prepare Your Opening Line</strong><br> I used to panic when someone asked, “So, what do you do?” Now I have a simple answer ready: “I work on AI systems that help insurance companies modernize their tech stack.” It is specific enough to be interesting, broad enough to start a conversation.</p><p>The truth is, everyone at that conference is just as nervous as you are. Most engineers are, to some extent, introverts. So if you stumble over your words or forget someone’s name, no one will judge you. They have been there too.</p><p>You are not alone in the crowd. You just need to find your voice.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b400ac696c19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | AI in Retail Inventory management]]></title>
            <link>https://medium.com/@pratyosh.desaraju/ai-in-retail-inventory-management-9f74be2d117e?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f74be2d117e</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Wed, 10 Apr 2024 12:00:05 GMT</pubDate>
            <atom:updated>2026-10-09T18:47:07.856Z</atom:updated>
            <content:encoded><![CDATA[<p>If you have ever walked into a massive warehouse store like The Home Depot or doom scrolled through Amazon, you have witnessed a complex action of logistics that most of us take for granted. We see a shelf with a product; we buy it. But behind that single entity lies an invisible nervous system managing millions of Stock Keeping Units (SKUs) across hundreds of locations spanning thousands of miles.</p><p>Retail inventory management is not just about counting boxes. It is a data problem.</p><p>For Mom and Pop stores down the street, managing inventory is extremely time consuming. For giants like The Home Depot, Walmart, Amazon the challenge is exponential. When we talk about enterprise retail, we are not talking about a few hundred items. We are talking about millions of active SKUs moving through a supply chain that never sleeps. They do not just need to know what they have; they need to know where it is, when it will run out, and how to get more there before the customer even realizes they need it.</p><p>For years, these systems were built on robust and rigid logic. They were designed to cater to the specific needs of each location based on historical averages.</p><p>Lets look at a warm and cold items scenario:</p><p><strong>The “Cold” Items:</strong> If a store is in Minnesota, the system knows that starting in October, sales for snow blowers and deicing salt will spike. It triggers stock orders based on the calendar.</p><p><strong>The “Warm” Items:</strong> If a store is in Florida, the system suppresses winter gear and boosts patio furniture or gardening tools year round.</p><p>These systems worked because seasons generally repeat. But “generally” is not good enough when margins are razor thin. A warm winter in the North or a sudden freeze in the South can leave retailers with millions of dollars in dead stock or empty shelves during a demand spike.</p><p><strong>Enter the era of Artificial Intelligence</strong><br> This is where AI can fundamentally change the architecture of inventory management. We are moving from predicting (guessing based on the past data) to sensing (analyzing the present in real time).</p><p>AI models do not just look at “last year’s sales.” They ingest a set of variables that a human analyst or a simple rule based code could never synthesize in real time.</p><p>Not just “it is winter,” but “a blizzard is hitting this specific zip code in 48 hours.”</p><p>It can move beyond historical data and ingest social media activity in rea time. Is a specific brand of power tool trending on social media in a specific region?</p><p>Systemic delays can be accounted for. Is a Suez Canal like incident delaying shipments? The AI enabled system can reroute inventory from a nearby locations before the shortage hits.</p><p>By using Machine Learning (ML) algorithms, companies can achieve higher accuracy. The system learns that a 5 degree temperature drop in Atlanta triggers a different buying behavior than the same drop in Boston. It stops treating stores as generic “locations” and starts treating them as unique data ecosystems.</p><p><strong>The Bottom Line</strong><br>In the end, the goal of AI in retail is not just efficiency; it is availability. We have moved past the era of guessing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f74be2d117e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Mentorship Matters: Cultivating the Next Generation of Innovators]]></title>
            <link>https://medium.com/@pratyosh.desaraju/mentorship-matters-cultivating-the-next-generation-of-innovators-d58459d8d687?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/d58459d8d687</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Sun, 21 Jan 2024 12:00:15 GMT</pubDate>
            <atom:updated>2026-10-09T18:46:53.952Z</atom:updated>
            <content:encoded><![CDATA[<p>Five years ago, I thought mentorship was a one way street, a generous act where experienced engineers bestowed wisdom upon eager newcomers. <br>I could not have been more wrong.<br>The truth is far more interesting.</p><p>Mentorship is not charity. It is not a favor we do for the industry or a box we check on our annual personal goals. It is an exchange that keeps the entire ecosystem alive.<br>Every year, my company takes on a fresh set of interns. Without fail, they arrive wide eyed, armed with questions that range from the practical (“How do I set up my development environment?”) to the philosophical (“Should we even be building this feature?”). <br>And yes, they ask a lot of questions.</p><p>I joke with my team that we need these interns around to humble us with their knowledge of whatever JavaScript framework released three weeks ago or the latest AI model that apparently makes everything we built last quarter obsolete. They show up citing stuff I have not read and using tools I have not heard of.<br>But here is the thing: that is not a joke. That is the point.</p><p>When you mentor someone, you are forced to articulate things you have internalized so deeply you forgot they were not obvious. You realize that best practice you have sworn by for years? It has exceptions. That architectural decision you made on autopilot? It requires updating.<br>Our interns do not just humble us with their technical knowledge. They humble us with their questions.</p><p>The engineer’s dichotomy, that tension between imposter syndrome and hubris, becomes especially obvious when someone is watching you work, asking you to defend your choices, expecting you to have answers. You suddenly realize how much of what you do is auto pilot, how many of your decisions are “it depends,” how often the honest answer is “I am not sure”.</p><p>Mentorship is not just good for individual growth. It is critical for the industry’s lifecycle.<br>Technology does not advance through working in isolation. It advances through knowledge transfer, through the passing of both technical skills and contextual wisdom from one generation to the next. When we fail to mentor effectively, we create knowledge silos, we lose memory, we watch newcomers repeat mistakes we solved decades ago.<br>The insurance industry, is a perfect example. We are in the midst of a massive transformation, bringing AI and machine learning to systems that have operated largely unchanged for decades. The only way this transformation succeeds is if experienced engineers who understand the domain work alongside fresh minds who understand the technology.<br>When our interns ask why we built something a certain way, they are not just learning our rationale. They are pressure testing whether that rationale still holds.<br><br>Now more than ever, young engineers need our support, mentorship, and advocacy.<br>There is a particular anxiety rippling through early career developers right now: the looming fear that AI will render their newly acquired skills obsolete before they have even had a chance to use them. They watch models generate code in seconds that took them hours to learn, and they wonder: what is the point?<br>This fear is not entirely unfounded, but it misunderstands what engineering actually is.<br>Yes, AI can generate boilerplate. Yes, it can suggest solutions to common problems. But engineering is not just writing code. It is understanding systems, making tradeoffs, communicating with stakeholders, debugging the weird edge cases, knowing when to build and when to buy, recognizing which problems need solving and which do not.<br>These skills require experience, context, and mentorship.<br><br>So yes, our interns ask exhausting numbers of questions. They challenge our assumptions, expose our knowledge gaps, and force us to justify our engineering decisions.<br><br>Mentorship matters, not as an act of charity, but as an investment in the future of the field we have chosen. The next generation of innovators is not just waiting to be taught. They are waiting to teach us, to challenge us, to push us forward.<br>And if we are lucky, they will keep us humble enough to keep learning, even after we think we have figured it all out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d58459d8d687" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Beyond the Hype: The AI Optimization Playbook]]></title>
            <link>https://medium.com/@pratyosh.desaraju/beyond-the-hype-the-ai-optimization-playbook-327b417cd462?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/327b417cd462</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Fri, 10 Nov 2023 12:00:41 GMT</pubDate>
            <atom:updated>2026-10-09T18:46:39.903Z</atom:updated>
            <content:encoded><![CDATA[<p>ChatGPT crossed 100 million users faster than any application in history. Today every earnings call mentions artificial intelligence. Venture capital is flooding into generative AI startups at a pace that makes even seasoned investors nervous. We are twelve months into what feels like a technological revolution, and the pressure to “do something with AI” has never been more intense.<br>But here is what the early data is starting to whisper, even as the hype machine roars: most organizations are struggling to explain what they are actually trying to accomplish. They know they need an AI strategy because their competitors have one. What they do not know is how to translate that power into measurable business outcomes.</p><p><strong>The Warning Signs</strong><br>If you look past the success stories and vendor case studies, a different pattern emerges. Historical data on AI and machine learning projects shows that 70 to 85% fail to deliver expected outcomes. Which means we are rushing headlong into technologies we have not yet learned to deploy successfully.<br>The challenge is not technological capability. GPT 4 can write code, analyze documents, and generate creative content with genuinely impressive fluency. The challenge is organizational readiness. Most companies are trying to bolt generative AI onto processes, systems, and cultures that were not designed for it.<br>Proving AI’s usefulness emerged as the greatest challenge in implementing AI initiatives in 2023. Not model performance. Not compute costs. Not regulatory concerns.</p><p>Usefulness.</p><p>That single word should give every executive pause, because it suggests that we are deploying solutions before we have clearly defined the problems.</p><p><strong>Data Quality</strong><br>Generative AI models are trained on vast datasets scraped from the internet. They perform remarkably well on general tasks because their training data is broad and diverse. But enterprise use cases require specificity, accuracy, and integration with proprietary systems and data.<br>Most organizations do not have their data in order. Documents are scattered across SharePoint, Google Drive, and local hard drives. Business processes are documented inconsistently or not at all. Critical knowledge lives in the heads of employees nearing retirement. You cannot feed chaos into an AI system and expect clarity to emerge.</p><p><strong>The Playbook</strong><br>The organizations that will succeed with generative AI are not the ones moving fastest. They are the ones moving most deliberately. Here is what that looks like in practice:<br>Define Success Before Deployment: Do not launch a generative AI pilot to “explore the technology” or “see what it can do.” Identify a specific business process, measure current performance, and establish clear metrics for what improvement looks like. Attempt to articulate the success criteria in one sentence.<br>Augmentation not Transformation: The highest value, lowest risk generative AI applications augment existing workflows rather than replacing them entirely. Document summarization. Meeting notes. Email drafting. Research synthesis. These use cases provide immediate value.<br>Invest in Data Readiness: Before spending another dollar on AI tools, audit your data infrastructure. Where is critical information stored? How consistent are formats and naming conventions? What gaps exist in documentation? The companies that solve these foundational issues first will extract far more value from generative AI than those that jump straight to implementation.</p><p><strong>The Path Through the Hype</strong><br>Generative AI is real. The capabilities are genuine. The potential applications span virtually every business function and industry. But potential and results are not the same thing. The gap between them is filled with strategy, execution, data quality, organizational change, and disciplined measurement.<br>As we close out 2023 and look toward 2024, the organizations that treat generative AI as an optimization challenge rather than a transformation mandate will be the ones that actually deliver results. They will start small, learn quickly, and scale what works. They will resist the pressure to deploy AI everywhere and instead focus on deploying it where it matters.<br>The hype cycle will continue. But the organizations that succeed will be the ones that tune out the noise and focus on the fundamentals: clear objectives, clean data, rigorous measurement, and incremental scaling.<br>It is not glamorous. But it may work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=327b417cd462" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | AI’s Promise for Insurance]]></title>
            <link>https://medium.com/@pratyosh.desaraju/ais-promise-for-insurance-20e061dc2c0a?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/20e061dc2c0a</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Thu, 10 Aug 2023 12:00:19 GMT</pubDate>
            <atom:updated>2026-10-09T18:46:22.120Z</atom:updated>
            <content:encoded><![CDATA[<p>The insurance industry has a secret that is barely a secret anymore: We are running a 21st century business on 20th century infrastructure, sometimes even 19th century thinking. While companies born in the smartphone era glide on cloud native architectures and real time data pipelines, those of us in insurance often find ourselves wrestling with codebases and mainframes that predate the now ubiquitous internet itself.<br>It is not for lack of trying, and it is certainly not for lack of intelligence. The issue is far more fundamental: insurance is old. Really old. <br>Over a century of accumulated systems, processes, and regulatory scaffolding. Compare that to the food delivery apps and ride sharing platforms that burst onto the scene in the last fifteen years. They had the luxury of building from scratch, architecting for the cloud, designing for mobile first experiences from day one.</p><p><strong>The Weight of Legacy</strong><br>Here is the uncomfortable truth: upgrading legacy technology in insurance is not just expensive, it is existentially risky. When your core systems process millions of policies, power billions of dollars in claims, and operate under strict regulatory oversight, you cannot simply flip a switch and migrate to the latest platform. The cost is not measured only in dollars (though those numbers can be staggering), but in operational risk, regulatory compliance, and the very real possibility of catastrophic failure.<br>I have watched teams spend years planning migrations that other industries complete in months. The difference? Insurance carries the weight of its own history: data models designed in the 1990s, business rules embedded in undocumented code, and integration points that snake through systems like archaeological layers.</p><p><strong>The Great Leveler</strong><br>Yet within this challenge lies something unexpected: opportunity. <br>Artificial intelligence, or more precisely, machine learning and intelligent automation, may be the field leveler insurance has been waiting for decades to find.​<br>Unlike traditional modernization efforts that demand rip and replace transformations, AI can work alongside legacy systems. It does not require you to rebuild everything from scratch. Instead, it augments what exists, filling gaps, automating manual processes, and extracting insights from data that has been sitting dormant in decades old databases.</p><p><strong>The Ways Forward</strong><br>The practical applications are not theoretical anymore. They are happening now, and they are reshaping how insurers operate:<br>Predictive Risk Modeling: Machine learning algorithms can analyze historical claims data, weather patterns, demographic shifts, and even social media signals to identify high risk areas before losses occur. Instead of relying solely on actuarial tables built on past performance, insurers can now anticipate emerging risks from climate change impacts to shifting fraud patterns with remarkable precision.<br>Claims Processing Automation: Natural language processing and computer vision can review claims documents, assess damage from photos, and flag anomalies for human review. What once took days of manual review can now happen in minutes, freeing adjusters to focus on complex cases that truly need human judgment. The efficiency gains are measurable, but the real win is improved customer experience which include faster settlements, less frustration, and more transparent processes.<br>Dynamic Pricing and Personalization: AI enables insurers to move beyond broad risk pools toward truly individualized pricing. By analyzing driving behavior, health metrics, property conditions, and countless other data points in real time, insurers can offer products that feel custom built rather than one size fits all. This is not just better business, it is fundamentally fairer, rewarding lower risk behaviors with lower premiums.</p><p>The transformation will not happen overnight, and it should not. Insurance serves a critical social function. It is the financial bedrock that lets people buy homes, start businesses, and recover from disasters. Recklessness in pursuit of innovation would be worse than stagnation.<br>But the tools are here. The technology works. And for an industry that has spent decades watching other sectors leapfrog ahead on the backs of better technology, AI represents something rare: a second chance. Not to abandon what we have built, but to augment it, enhance it, and carry it forward into a future where legacy does not mean obsolete, it means optimized.<br>The question is not whether AI will transform insurance. It is whether we will seize the opportunity to guide that transformation, or whether we will let it happen to us. The companies that choose the former will be the ones still standing when the next century of insurance history gets written.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20e061dc2c0a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Artificial Intelligence 101]]></title>
            <link>https://medium.com/@pratyosh.desaraju/artificial-intelligence-101-766cd64e91dc?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/766cd64e91dc</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Wed, 28 Dec 2022 12:00:28 GMT</pubDate>
            <atom:updated>2026-10-09T18:46:08.356Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial Intelligence isn’t new. The term was coined in the 1950s yet only recently has hype yielded real results. We have moved past the AI winter into what feels like spring with important caveats. The field has matured enough to deploy AI systems in production yet remains nascent enough that every implementation still teaches fundamental lessons about both the technology.</p><p>The AI versus ML Distinction:</p><p>Artificial Intelligence is the broad field focused on creating machines that mimic human cognitive functions such as reasoning and learning.</p><p>Machine Learning is a subset of AI in which we supply algorithms with data so they can learn patterns without being explicitly programmed for every scenario.</p><p>Most systems called AI today rely on Machine Learning rather than any general intelligence. For example when Netflix suggests a movie or Gmail filters spam these are Machine Learning systems not true artificial intelligence.</p><p>Five years ago the joke doing the rounds was “AI is just a bunch of if conditions.” The joke may no longer be funny because we have demonstrable products.</p><p>So what is next? That part we don’t know for sure.</p><p>Some CEOs will claim we don’t need pilots or programmers anymore since AI is here. Naysayers will say AI is not there yet. Which view is true? Time will tell. One thing AI has going for it is that it is data driven so progress will likely exceed linear growth perhaps even approach exponential.</p><p>The possibilities are really cool though. Image generation is one popular use being touted. Maybe video generation is next?</p><p>The bigger challenge for AI will be making inference cheaper. Mass adoption will only happen once that is achieved. Think of how low data prices sparked the smartphone era.</p><p>One thing is certain: changes are coming. Quick.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=766cd64e91dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Extreme Programming: When Your Team Finally Gets It]]></title>
            <link>https://medium.com/@pratyosh.desaraju/extreme-programming-when-your-team-finally-gets-it-1146316309d0?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/1146316309d0</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Thu, 10 Feb 2022 12:00:08 GMT</pubDate>
            <atom:updated>2026-10-09T18:45:54.510Z</atom:updated>
            <content:encoded><![CDATA[<p>I remember the first time I suggested pair programming to my team. The reaction was immediate and predictable. “That’s inefficient”, “Two people doing the job of one?” , “We’ll get half as much done.”</p><p>We weren’t wrong in theory. But we were wrong in practice.</p><p><strong>The Pair Programming Resistance<br></strong>Pair programming sounds counterintuitive. Two engineers at one keyboard. Surely that’s a waste of resources. Surely you could get more done if they worked separately. That’s what I thought too, until I actually tried it.</p><p>What changed my mind wasn’t ideology. It was watching my team discover something unexpected. After a few weeks of pairing, knowledge started flowing. The junior engineer learned patterns from the senior. The senior engineer got fresh perspectives from the junior. Bugs that would have taken hours to debug alone were caught instantly. Code reviews became less time consuming because the code was already reviewed as it was written.</p><p>Within three months, the same team that resisted pairing realized they were shipping better code faster. Not despite the pairing. Because of it.</p><p>The only problem we discovered was that we always needed an even number of engineers. We joked about it constantly. Especially before I went skydiving.</p><p><strong>Test Driven Development<br></strong>Once the team saw the value of collaboration through pairing, TDD became easier to sell. Write tests first. Then write code to pass those tests. It sounds backward until you realize you’re actually designing your code before you build it.</p><p>TDD forces you to think about what you’re building before you build it. Your tests become specifications. Your code becomes an implementation. Bugs that would have lived in production for months get caught immediately because the tests reveal them.</p><p><strong>Behavior Driven Development<br></strong>BDD takes TDD further. Instead of writing unit tests in developer speak, you write tests that describe behavior in business terms. Given this scenario, when this happens, then this should result. Suddenly your tests read like requirements.</p><p>I’ve seen BDD transform conversations between engineers and product managers. When you write tests in business language, product people understand exactly what the code is supposed to do. They can read the tests and verify the behavior matches expectations. No more miscommunication. No more building the wrong thing.</p><p><strong>Continuous Integration<br></strong>Pair programming, TDD, and BDD only work if you’re integrating constantly. Every commit goes into the main branch. Every commit runs the full test suite. If something breaks, you know immediately. You fix it immediately.</p><p>No more integration nightmares where you branch for two weeks and come back to merge conflicts and broken code. CI forces discipline. It forces you to think about what you’re shipping.</p><p><strong>The Transformation<br></strong>These practices seem extreme at first. Pair on everything? Write tests before code? Think about behavior before implementation? Integrate multiple times a day?</p><p>But they’re not extreme. They’re just systematic approaches to building better software. My team discovered that individually.</p><p>We were averse to all of it initially. Now we won’t go back. Knowledge sharing through pairing means less onboarding time for new engineers. TDD means fewer production bugs. BDD means better alignment with product. CI means confidence in every deploy.</p><p>Extreme programming isn’t extreme when you see the results. It just becomes how you work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1146316309d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | Testing: The ZOMBIE strategy]]></title>
            <link>https://medium.com/@pratyosh.desaraju/testing-the-zombie-strategy-b50c0edb1e17?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/b50c0edb1e17</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Sun, 02 May 2021 12:00:37 GMT</pubDate>
            <atom:updated>2026-10-09T18:45:37.195Z</atom:updated>
            <content:encoded><![CDATA[<p>Testing can feel overwhelming. You know you need tests, but where do you even start? Do you test everything? Do you focus on coverage percentages? Most engineers end up writing random tests that technically pass but don’t actually protect the code.</p><p>That’s where ZOMBIE comes in. It’s a simple framework that gives you a systematic approach to testing.</p><p><strong>Z is for Zero</strong></p><p>Start with nothing. Test your code when you give it zero inputs or initial conditions. What happens when you create an object with no data? What are the defaults? This catches initialization bugs before anything else builds on top of them.</p><p><strong>O is for One</strong></p><p>Test with a single or minimum valid input. The simplest case. Does your code handle it correctly? We often skip this because it seems too obvious, but that’s exactly where bugs hide.</p><p><strong>M is for Many</strong></p><p>Now scale it up. Test with multiple inputs, complex scenarios, larger datasets. Does your code still work? Does it break under load? Can it handle what production will actually throw at it?</p><p><strong>B is for Boundary</strong></p><p>Test the edges. Zero, negative numbers, maximum values, empty collections. Boundary conditions are where most bugs live because developers rarely think about them.</p><p><strong>I is for Interface</strong></p><p>Before you worry about how it works inside, make sure the interface itself is good. Is it easy to use correctly? Is it hard to use incorrectly? A confusing interface creates problems regardless of implementation quality.</p><p><strong>E is for Exception</strong></p><p>Test failures. Invalid inputs. Missing data. Null values. What happens when things go wrong? Does your code fail gracefully or crash?</p><p><strong>S is for Simple</strong></p><p>Don’t overcomplicate your tests. Keep scenarios straightforward. Keep solutions simple. Complex tests break easily and become hard to maintain.</p><p><strong>Why It Works</strong></p><p>ZOMBIE works because it’s systematic. You’re not testing randomly or guessing what matters. You’re following a framework that naturally covers the scenarios where real bugs happen.</p><p>Use ZOMBIE. Your tests will be better, and you’ll actually catch bugs before they reach production.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b50c0edb1e17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pratyosh Desaraju | The engineer’s dichotomy: Imposter and Hubris]]></title>
            <link>https://medium.com/@pratyosh.desaraju/the-engineers-dichotomy-imposter-and-hubris-b2d5a9fb6b2f?source=rss-e0ec3481c855------2</link>
            <guid isPermaLink="false">https://medium.com/p/b2d5a9fb6b2f</guid>
            <dc:creator><![CDATA[Pratyosh Desaraju]]></dc:creator>
            <pubDate>Sat, 05 Dec 2020 00:23:15 GMT</pubDate>
            <atom:updated>2026-10-09T18:44:41.608Z</atom:updated>
            <content:encoded><![CDATA[<p>We all know the story of Icarus:</p><p>Imprisoned by the King on an island south of the Aegean Sea, his father Daedalus devised an escape plan. He planned to gather feathers from the birds that flew near their tower, secure them with thread, and fasten them with wax. He crafted two pairs of wings, one for himself and one for his son.</p><p><em>The Imposter:</em><br>Yet, Daedalus questioned his own design. He wallowed in self doubt for months. Would the wings hold? Was he leading himself and Icarus into disaster? Their escape plans were delayed due to the weight of uncertainty.</p><p>Before they took flight, Daedalus gave Icarus strict instructions: fly neither too low, where the sea’s dampness could clog the wings, nor too high, where the sun’s heat could melt the wax.</p><p><em>The Hubris:</em><br>Icarus had other ideas. As he took his first steps into the air, he felt a sense of freedom. With each flap of the wings, he soared faster. In his excitement, he ignored Daedalus’ only warning and flew higher. The wax that bound the feathers began to melt. One feather came undone, followed by another. Soon, he was in free fall, and Daedalus watched in horror as his own child plummeted from the sky. All it took was a moment of hubris to bring about tragedy.</p><p>.<br>.<br>.</p><p>Cut to real world 3000 years later. A software engineer working in the heart of Silicon Valley experiences similar emotional swings every week.</p><p>One day, you can’t figure out why your Spring Boot application won’t start. You’ve checked the code four times, and a teammate joins you in reviewing it. Still, no luck. Four hours later, after nearly emailing your bootcamp demanding a full refund and an unconditional apology, you find it: an em dash sneaked into your configuration file name where a regular dash should be.</p><pre>application–dev.properties</pre><p>looks identical to</p><pre>application—dev.properties</pre><p>At least on the screen it does, but Spring doesn’t recognize the file name with the em dash character. Your caffeine addled brain didn’t catch the Unicode difference between (U+2014) and (U+002D) either.</p><p>You hate yourself. You feel like the worst software engineer since the developer who hard coded their password into the source code.</p><p>The next day morning, you log in. The walk of shame to the daily stand up, where you tell everyone you spent the entire day fixing your own damn configuration file, beats down your confidence.</p><p>Moments later: The dreaded Microsoft Teams chime rings across the floor…</p><p>Users cannot log in across the country. The manager and product owner have spun up a call. Teammates join in, and so do you. No one can figure out what’s going on while looking at the dashboard.</p><p>Different theories are thrown around. Some suggest restarting the containers. It doesn’t help.</p><p>The senior developer is going through recent commits. The junior developer repeats whatever the managers say, just to be heard on the call.</p><p>You notice something.</p><p>The number of calls coming into your service and the number of errors logged don’t line up. Suddenly, you see it. A call to a vendor’s service times out, and the code doesn’t account for it. The application attempts to carry on like nothing happened when it should have errored out and notified the team.</p><p>You have a fix! The vendor has been informed of the timeout. Their containers are restarted. Users can login again.</p><p>Cue Borat’s great success gif!</p><p>Yesterday, you couldn’t look yourself in the eye, and today you’re the manager’s favorite (at least for today; we’ll see how tomorrow goes). This is the see saw of emotions of the engineering world.</p><p>How do we deal with this?</p><p>The key is to embrace this dichotomy: not as a binary choice but as a dynamic range of experiences. Every line of code, every bug, and every fix contributes to a broader journey of learning and growth. Instead of viewing yourself as an imposter or a superior developer, try to see yourself as a perpetual learner. Daedalus acknowledged his limitations while continually striving to innovate. The skies may be unpredictable, but with resilience and awareness, he knew he could soar with purpose.</p><p>In the world of Icarus, be a Daedalus.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b2d5a9fb6b2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>