<?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 Puneet Udhayan on Medium]]></title>
        <description><![CDATA[Stories by Puneet Udhayan on Medium]]></description>
        <link>https://medium.com/@puneetudhayan?source=rss-67c0a64a3979------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*QDN5HnCpi6Q6q7cpc9JjuQ@2x.jpeg</url>
            <title>Stories by Puneet Udhayan on Medium</title>
            <link>https://medium.com/@puneetudhayan?source=rss-67c0a64a3979------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 12:27:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@puneetudhayan/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[From Email to Empty Plate: How I Built an AI-Powered ‘Smart Fridge’ to Minimize Food Waste with No…]]></title>
            <link>https://medium.com/@puneetudhayan/from-email-to-empty-plate-how-i-built-an-ai-powered-smart-fridge-to-minimize-food-waste-with-no-f3a3519c031a?source=rss-67c0a64a3979------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3a3519c031a</guid>
            <category><![CDATA[n8n]]></category>
            <category><![CDATA[no-code]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[food-waste]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Puneet Udhayan]]></dc:creator>
            <pubDate>Tue, 30 Sep 2025 02:32:09 GMT</pubDate>
            <atom:updated>2025-09-30T02:32:09.312Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Email to Empty Plate: How I Built an AI-Powered ‘Smart Fridge’ to Minimize Food Waste with No Code</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uOxv0oNTSAEBsZQNsEjXlQ.jpeg" /><figcaption>Photo by Ron Lach : <a href="https://www.pexels.com/photo/opened-fridge-food-waste-symbol-8466649/">https://www.pexels.com/photo/opened-fridge-food-waste-symbol-8466649/</a></figcaption></figure><p>I have a confession: I waste a lot of food. It’s a cycle of guilt. I buy fresh ingredients for a recipe, use half of them, and then the rest wither away in the back of my fridge. Leftovers from last night? Forgotten. That half-used jar of sauce? A future science experiment.</p><p>I knew I needed a system — a kind of fridge database — to remind me what I have and when I need to use it. But as a developer, I also know my own worst enemy: the desire to build the “perfect” app from scratch. I’ve started countless elaborate projects with all the bells and whistles, only to abandon them before they’re ever finished.</p><p>This time, I decided to do things differently. My goal was <strong>speed and immediate value</strong>. I challenged myself to build a working solution without writing a single line of code.</p><h3>The Game Plan: No-Code Automation with n8n</h3><p>I decided to use <strong>n8n</strong>, a low-code automation platform. My plan was to intercept the grocery and food delivery receipts that already land in my inbox and use them to automatically build my food inventory.</p><p>I was shocked by how quickly it came together. In just a couple of hours — including the time it took to learn the basics of n8n — I had a fully functional prototype. The platform was very intuitive. I only needed to reference the document once to set it up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_4CqQToooAJWUP7THuQw-A.png" /><figcaption>screenshot of the n8n pipeline</figcaption></figure><h3>How the “Smart Fridge” Workflow Works</h3><p>The entire process is a simple, automated workflow that runs three times a day, about an hour before my usual meal times.</p><ol><li><strong>Read Emails:</strong> The workflow starts with a <strong>Gmail Trigger node</strong>. It scans my inbox for new mail.</li><li><strong>Filter the Noise:</strong> A <strong>Filter node</strong> immediately discards irrelevant emails, only keeping receipts from my favorite food delivery and quick commerce apps.</li><li><strong>The AI Magic Extractor:</strong> This is where it gets really cool. I used an AI-powered node to extract information from the email’s HTML body. I simply gave it a sample JSON format like {&quot;items&quot;: [{&quot;name&quot;: &quot;...&quot;, &quot;quantity&quot;: &quot;...&quot;}]}, and the LLM (OpenAI&#39;s GPT-4.1-mini) did the rest. It intelligently read the email and pulled out exactly what I bought. This is a game-changer! No more writing custom web scrapers for each company that break the moment they update their email template.</li><li><strong>Estimate Shelf Life:</strong> The list of items is then passed to another AI node (<strong>Basic LLM Chain</strong>). I prompted a more powerful model (GPT-4o) to estimate the shelf life for each grocery item. I’ll admit, this part is still a work in progress. The AI’s estimates were occasionally off.</li><li><strong>Store and Notify:</strong> The items and their “use by” dates are added to a database. The workflow then queries the database, grabs the top 10 items closest to expiring (ignoring anything already past its date), and sends me a neat summary email titled “What to Eat Next!”</li></ol><h3>Early Results and What’s Next</h3><p>This quick-and-dirty MVP already works! I’m getting timely reminders that are helping me use up ingredients before they go bad.</p><p>But this is just the beginning. I’m already planning several improvements:</p><ul><li><strong>Better Shelf-Life Logic:</strong> Instead of relying on the LLM for shelf life, I’m going to cluster items into predefined categories (e.g., “Leafy Greens,” “Dairy,” “Meat”) with manually set, reliable expiration periods.</li><li><strong>Smart Recipe Suggestions:</strong> The next logical step is to have the system suggest recipes that use the ingredients that are about to expire.</li><li><strong>A Simple Mobile App:</strong> I’m thinking of building a simple companion app to let me quickly mark items as “consumed” or “cooked,” which will help keep my fridge database perfectly in sync with reality.</li></ul><p>This project was a powerful reminder that you don’t always need to spend weeks coding to build something incredibly useful. With modern tools like n8n and the power of LLMs, a great idea can become a working reality in a single afternoon.</p><p>Want to try it for yourself? Here is a link to my n8n pipeline:<br><a href="https://github.com/PuneetUdhayan/fridge-db/blob/main/n8n_pipelines/Fridge%20DB.json">https://github.com/PuneetUdhayan/fridge-db/blob/main/n8n_pipelines/Fridge%20DB.json</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3a3519c031a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An approach to secure Long-Running Jobs in a Microservices Architecture]]></title>
            <link>https://medium.com/@puneetudhayan/an-approach-to-secure-long-running-jobs-in-a-microservices-architecture-b8895af87b49?source=rss-67c0a64a3979------2</link>
            <guid isPermaLink="false">https://medium.com/p/b8895af87b49</guid>
            <dc:creator><![CDATA[Puneet Udhayan]]></dc:creator>
            <pubDate>Wed, 20 Aug 2025 01:55:12 GMT</pubDate>
            <atom:updated>2025-08-20T01:55:12.089Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oS0xaRLbqLrsQPJnBKMaaQ.png" /></figure><p>In a modern microservices environment, authentication is the bedrock of security. For user-facing requests, the flow is straightforward: a user logs in, receives a bearer token, and every subsequent API call is authenticated with that token.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YAljzfHv76d_-dOuVH24-w.png" /><figcaption>Synchonous communication in microsevices</figcaption></figure><p>But what happens when the work isn’t immediate? What about the long-running, asynchronous jobs that are crucial to our platform — things like generating complex reports, processing large datasets, or transcoding media files? User tokens expire, and for good reason. But that very feature makes them unsuitable for tasks that might run for hours or even days.</p><h4>The Initial Problem: The Limits of Simple Authentication</h4><p>The most obvious solution to authenticate a long-running job is to use a static service ID that represents the service. The service can obtain a token for this service ID, which can be used for making requests. Since this is a fixed service ID, all other services will know that the request originated from a particular service.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9PEZtyJpY-xiZH0eoWNHkg.png" /><figcaption>Allow list service ID flow</figcaption></figure><p>While this works, it presents a significant security risk.</p><p>If that service token were ever leaked, it would become a powerful, persistent “master key.” An attacker could use it to impersonate the service indefinitely, potentially accessing or manipulating any data the service was authorized to touch. This approach violates a core security tenet: the <strong>Principle of Least Privilege</strong>, which dictates that any entity should only have the bare minimum permissions necessary to perform its function.</p><h4>Solution: The “Task Credentials” System</h4><p>To solve this, IBM uses a system that delegates user authority securely without relying on long-lived, high-privilege tokens. The Task Credentials system ensures that long-running jobs still operate on behalf of the user, but in a way that is secure, auditable, and temporary.</p><p>Here’s how it works, step by step:</p><ol><li><strong>Job Initiation:</strong> A user, authenticated with their standard bearer token, kicks off a long-running job through an API call.</li><li><strong>Credential Creation:</strong> The initial service receives this request. Before starting the asynchronous job, it uses the user’s valid token to make a call to our authentication service, generating a new, dedicated API key specifically for this task. This new key inherits the necessary permissions from the user but is scoped only to this job.</li><li><strong>Secure Storage:</strong> This newly created, task-specific API key is immediately stored in a secure, centralized vault (like HashiCorp Vault or AWS Secrets Manager). It is never passed around directly.</li><li><strong>Delegated Retrieval:</strong> The asynchronous worker service that needs to perform the job uses its own, low-privilege service ID to authenticate with a dedicated microservice called <strong>Task Credentials</strong>.</li><li><strong>On-Demand Access:</strong> The Task Credentials service validates that the worker is authorized to act on this specific job. If the request is valid, it fetches the temporary API key from the vault and returns it to the worker.</li><li><strong>Secure Execution:</strong> The worker service now uses this temporary API key to make authenticated calls to other microservices. It is now effectively and securely acting on behalf of the original user. Once the job is complete, the API key can be revoked.</li></ol><p>This flow elegantly solves the problem. The powerful credential (the user-delegated API key) is short-lived and tightly controlled, while the static credential (the worker’s service ID) has almost no power of its own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b6IA63awoJ_b7jlUPswiVg.png" /><figcaption>Task credentials flow</figcaption></figure><h4>But Isn’t the Service ID Still a Risk?</h4><p>This is the crucial question. If the worker service still uses a service ID, haven’t we just moved the problem?</p><p>The answer is no, because we have drastically <strong>reduced the scope of the credential</strong>. The service ID used in this new system has only one permission: to ask the Task Credentials service for a job key.</p><ul><li><strong>Old Way:</strong> A leaked token was a <strong>master key</strong> that could open any door the service had access to. The blast radius of a compromise was enormous.</li><li><strong>New Way:</strong> A leaked token is like a <strong>request form</strong>. All an attacker can do with it is ask the Task Credentials service for a key. This request is heavily audited, and the service will only hand over a key for a legitimate, active job. The potential for misuse is dramatically minimized.</li></ul><p><strong>Why Bother? The Case for Internal Authentication</strong></p><p>This entire system is built on a foundational security concept: <strong>Zero Trust</strong>. We assume that no request is inherently trustworthy, even if it originates from within our own network. Authenticating every single call between our microservices is non-negotiable for three reasons:</p><ol><li><strong>Authorization:</strong> Authentication answers, “Who are you?” Authorization answers, “What are you allowed to do?” Without knowing who is making a request, we can’t enforce critical access rules.</li><li><strong>Defense in Depth:</strong> If an attacker breaches our network perimeter, we don’t want them to have free rein. Inter-service authentication acts as a series of internal checkpoints, containing any potential breach and preventing an attacker from moving laterally through our system.</li><li><strong>Auditing and Accountability:</strong> When every action is tied to an authenticated service (and, in this case, a delegated user context), we get a crystal-clear audit trail. If something goes wrong, we know exactly who to blame, making debugging and security forensics infinitely easier.</li></ol><p>By moving from a simple but brittle authentication method to the Task Credentials system, we’ve built a more resilient, secure, and accountable platform — one that protects both our infrastructure and our users’ data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b8895af87b49" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[IBM watsonx.governance model untrack issue]]></title>
            <link>https://medium.com/@puneetudhayan/ibm-watsonx-governance-model-untrack-issue-674064717160?source=rss-67c0a64a3979------2</link>
            <guid isPermaLink="false">https://medium.com/p/674064717160</guid>
            <category><![CDATA[watsonx-governance]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Puneet Udhayan]]></dc:creator>
            <pubDate>Thu, 24 Jul 2025 04:53:05 GMT</pubDate>
            <atom:updated>2025-07-24T04:53:05.658Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*edjJ0N7oxZhUUKzgefKavQ.png" /></figure><h3>Issue Summary</h3><p>When promoting a model into a space or project (via export and import), the model may automatically untrack.</p><h3>Cause</h3><p>This issue may occur if your access to the AI use case is managed through an access group rather than direct permissions. In such cases, a required service ID might lack the necessary access to track models within the target space or project.</p><h3>Resolution</h3><p>To resolve this issue, we can add the <strong>“All IAM account management services”</strong><em> </em>access policy<em> </em>to the required service ID.</p><h3>STEP 1</h3><p>Navigate to <a href="https://cloud.ibm.com/iam/serviceids">https://cloud.ibm.com/iam/serviceids</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d4Y3i3LxyYUYDXrKEAkeXA.png" /></figure><p>STEP 2</p><p>Search for <strong>watsonx.governance</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9Pxzxja_rww-_iW5l1FR7w.png" /></figure><p>STEP 3</p><p>Unlock the service ID with the name <strong>watsonx.governance_DO_NOT_DELETE_&lt;your-region&gt;</strong> by clicking on <strong>List of actions</strong>, then <strong>Unlock</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hc90Srq9Cjx3L0FToLKdTg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EUsQdg8NlnnqrQ8r6iMyxQ.png" /></figure><p>STEP 4</p><p>Click on the service ID’s name, which will take you to its landing page, then click on <strong>Assign access</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EAvqk7_Ntad3ygfNiQOpFQ.png" /></figure><p>STEP 5</p><p>In the service section, search for <strong>“All IAM account management services</strong>,” select it, and then click <strong>“Next</strong>.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FlITCcSEFKOlxFycDfAkiw.png" /></figure><p>STEP 6</p><p>For the “<strong>Roles and Actions”</strong> select the “<strong>Viewer”</strong> role under “<strong>Platform access”</strong>. Then click on <strong>“Next”</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MuzowK-eBdx51EAUdl_azw.png" /></figure><p>STEP 7</p><p>Click on “<strong>Add”</strong>. Then click on “<strong>Assign”</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QsqNtAeJZDukE01oLMjQAA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kSe7klZZSYMrEPvoAdBnxg.png" /></figure><p>You should now see “<strong>All IAM Account Management services”</strong> under the service IDs “<strong>Access policies”</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H0dfh9oURb1bR2r-QZLCNw.png" /></figure><p>STEP 8</p><p>Lock the service ID by navigating to the list of service IDs, finding the required service ID, and finally clicking on “<strong>Lock”</strong> from the “<strong>List of actions”</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V55tBEhxWsTIr0Dcnt0F5Q.png" /></figure><p>Hope this helps :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=674064717160" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Orchestrating Data Pipelines with Large Language Models]]></title>
            <link>https://medium.com/@puneetudhayan/orchestrating-data-pipelines-with-large-language-models-cc97b0123067?source=rss-67c0a64a3979------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc97b0123067</guid>
            <category><![CDATA[data-pipeline]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent-tutorial]]></category>
            <dc:creator><![CDATA[Puneet Udhayan]]></dc:creator>
            <pubDate>Thu, 03 Jul 2025 07:57:03 GMT</pubDate>
            <atom:updated>2025-07-03T07:57:03.679Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1013/1*z-Lz42vxltuXPE2-oMNrXA.png" /></figure><p>Before LLMs were popular, I worked on a project that involved deploying and scaling data scrapers. To get the results we needed, we ended up chaining two or more scrapers together. Based on the customer’s requirements, we found ourselves building custom pipelines with complex if-else logic and data mapping rules. This is where the transformative power of LLMs enters the picture.</p><p>This article presents a method for dynamically building pipelines based on user prompts and demonstrates it with an example. The approach is built on the ReAct framework with some key modifications.</p><p>Two Amazon scrapers have been built that will be provided as tools to the LLM. A “product listing scraper” (amazon_plp_scraper) and a “product page scraper”(amazon_product_scraper). The amazon_plp_scraper gets the URLs of all products from an Amazon product listing page. Below is an image of a product listing page.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XSPtML_NymtEVdIA" /><figcaption>Amazon product listing page</figcaption></figure><p>The Amazon product page scraper returns the details of a given product. Attached below is an image of a product page that the <em>amazon_product_scraper</em> accepts as input. Also provided is the input and output schemas of both tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oxMs-aL2EhdxymXa" /><figcaption>Amazon product page</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qqLAeKgVDG5SF98PtEIwzA.png" /></figure><p>Let us suppose the user gives us the prompt</p><blockquote><em>Can you get price rating and asin info from </em><a href="https://www.amazon.in/s?k=shampoo&amp;crid=VFUZU3H1K4BO&amp;sprefix=shampoo%2Caps%2C215&amp;ref=nb_sb_noss_2"><em>https://www.amazon.in/s?k=shampoo&amp;crid=VFUZU3H1K4BO&amp;sprefix=shampoo%2Caps%2C215&amp;ref=nb_sb_noss_2</em></a></blockquote><p>We will supply this prompt to an agent similar to the ReAct architecture as shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/616/0*uU7WydBaMu_q4LR9" /><figcaption>agent graph</figcaption></figure><p>The prompt is sent to the LLM, which is depicted as <em>tool_calling_llm</em> in the diagram. Here, the LLM looks at the URL and decides to call <em>amazon_plp_scraper</em>. This tool call information is sent to the <em>custom_tool_call</em> node. This is where we differ from the traditional ReAct architecture. Instead of executing the tool call, another LLM generates a dummy output for this tool call. We don’t want to execute the <em>amazon_plp_scraper</em> because it’s a long-running, resource-intensive job that generates a large volume of data. The goal is to build a pipeline, not execute this specific job, especially since we’re not yet sure if an output is possible with the given tools. This dummy output is sent back to the <em>tool_calling_llm</em>, which will now decide to call the <em>amazon_product_scraper</em>. Yet again, dummy output will be generated for this tool call and sent back to the tool_calling_llm.</p><p>At every step, we keep track of what tools have been called and with what inputs. We also track the dummy data generated by the custom_tool_call node. Using this information, we can piece together the order in which tools need to be called and the output-to-input mapping between the tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YrR8YLL_6mVVSN6O" /></figure><p>This information, which was prepared manually based on a customer’s needs, is now generated automatically by the LLM. You can see this in action using the notebook I created here:</p><p><a href="https://github.com/PuneetUdhayan/data-pipeline-llm/blob/main/DataPipelines.ipynb">https://github.com/PuneetUdhayan/data-pipeline-llm/blob/main/DataPipelines.ipynb</a></p><p>Stay tuned to see how this data can be used to execute scrapers in a scalable manner using the web queue worker architecture.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc97b0123067" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why write?]]></title>
            <link>https://medium.com/@puneetudhayan/why-write-a65a42748a70?source=rss-67c0a64a3979------2</link>
            <guid isPermaLink="false">https://medium.com/p/a65a42748a70</guid>
            <category><![CDATA[mediocrity]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Puneet Udhayan]]></dc:creator>
            <pubDate>Tue, 24 Jun 2025 04:10:22 GMT</pubDate>
            <atom:updated>2025-06-24T04:10:22.007Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fhWuEdOnEjC_40-K.jpg" /><figcaption>Hannah Cullwick</figcaption></figure><p>Let me take you back to London, 1851. One-third of all young women, those between the ages of fifteen and twenty-five, were servants. That’s an astonishing number of people whose lives went largely undocumented. We would know nothing about what they did at work, what they ate, or how they spent their time if it weren’t for one Hannah Cullwick.</p><p>Hannah Cullwick was born on 26th May 1833 and entered a life of servitude at the age of 8. From the age of twenty-five to her sixties, she kept very detailed diaries. The diaries mostly contained chores she had done on a given day. Here is an excerpt:</p><blockquote>Swept &amp; dusted the rooms &amp; the hall. Laid the hearth &amp; got breakfast up. Clean’d 2 pairs of boots. Made the bed and emptied the slops. Clean’d &amp; washed the breakfast things up.</blockquote><p>These mundane entries give us one of the only insights we have into the lives of a large section of Victorian women. Histories will always be written of Grand achievements and a few great outliers, but the human story, the stories of the majority of everyday people, are largely left out. Aren’t you curious about what your average Roman did when Ceaser ruled? What did they have for lunch? What did they aspire to become?</p><p>I’m sure there are more curious souls like me, and there will be more in the future. So, let’s write to them about the funny cat video we saw on our handheld communication device or about the new salad recipe we tried. Don’t worry about proper grammar or spelling errors; they only add character. So, pick up your pen, or more likely your phone, and start writing/typing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a65a42748a70" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>