<?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 Sillyfellow on Medium]]></title>
        <description><![CDATA[Stories by Sillyfellow on Medium]]></description>
        <link>https://medium.com/@sillyfellow?source=rss-557be62cc30b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*f1fvj2-KayKbW_i1HgcjfQ.jpeg</url>
            <title>Stories by Sillyfellow on Medium</title>
            <link>https://medium.com/@sillyfellow?source=rss-557be62cc30b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 03:45:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sillyfellow/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[Conversational Finance: AI Assistant that talks to your Fund Data]]></title>
            <link>https://medium.com/kotaicode/conversational-finance-ai-assistant-that-talks-to-your-fund-data-8f644336ab1f?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8f644336ab1f</guid>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[vc-funds]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Mon, 24 Nov 2025 07:13:47 GMT</pubDate>
            <atom:updated>2025-11-24T15:19:56.796Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XUJodcQWGtnH9ELbZ4sazA.png" /><figcaption>Talk to your fund.</figcaption></figure><p>What you see in the image is absolutely real, and for a real VC fund*.</p><p>We have been working with VC funds and taxation for sometime, and we thought it is high time we had a new way for fund managers and tax specialists to interact with financial data — naturally, securely, and instantly.</p><p>One of the key things is <strong><em>securely</em></strong> — and that’s where we have put MCP to good use. But, before going deeper into the hairy tech details, let me paint the whole picture.</p><h3><strong>The reality — everyone wants to know the numbers</strong></h3><p>Be it the quarter-end, or year end, the partners want to know the details of taxable income (among other things).</p><p>In Talix, you can login, navigate to the correct screen/s and then get the numbers for yourself. It is actually a breeze. We usually have someone lookup, and report. I don’t think we need to anymore.</p><p>Talix now includes a conversational AI assistant that lets general partners, and tax professionals interact with their financial data through natural language. You ask questions the way you’d ask a knowledgeable colleague, and the system responds with relevant information — while ensuring you only see data you’re authorized to access.</p><p>For an industry that has relied on spreadsheets, static reports, and complex navigation hierarchies for decades, this is big. I guess it is worth a few millions in funding (if I am going with the trend of AI bubble(?)).</p><h3><strong>Why This Matters (optional)</strong></h3><p>VC Fund admin/management/etc. is hard. Too many things to look up, and too many people to answer to. The challenge isn’t just the volume of data — it’s the friction involved in accessing, and presenting it.</p><p>Traditional software interfaces require navigation. And since you use it only when you’re in a hurry, they drive you mad.</p><p>Talix has a brilliant interface (this could be some kind of Ikea effect, as I did contribute to it), still it would be nicer to circumvent all that.</p><p>So, we opened it up with TalixGPT — carefully because of the significance of data security. A tax specialist managing multiple funds should see all of them. A general partner should only see funds they’re involved with. An AI assistant that ignores these boundaries would be worse than useless.</p><h3><strong>Quick Overview</strong></h3><p>In Talix, authorized users — currently tax specialists and general partners — see a chat button that opens a conversation window.</p><p>The assistant understands context. It can search for funds by name, retrieve detailed financial statements, and explain what data it found.</p><p>While respecting existing permissions at every step. When a general partner asks to see their funds, they see only their funds — the same data they’d access through the traditional interface, just delivered through a more natural interaction.</p><p>The assistant handles ambiguity gracefully. If it can’t find an exact match for a fund name, it offers suggestions. If you ask for something you don’t have access to, it explains that politely rather than failing silently.</p><p>From a user’s perspective, it is having an assistant who has all the spreadsheets at hand, is always available, and never puts you on hold (except for the time they need to lookup the numbers).</p><h3><strong>Security and Trust: How we protect the data</strong></h3><p>In everything with financial workflows, security is (and must be) the foundation. We designed the system with three distinct layers of protection, each serving a specific purpose.</p><h4><strong><em>Authentication First</em></strong></h4><p>Every chat request requires a valid login. The system verifies your identity before processing any question — the very same authentication that protects the rest of the application.</p><h4><strong><em>Role-Based Access</em></strong></h4><p>We made a deliberate choice to enable it only for tax specialists and general partners initially. This isn’t a technical limitation — it’s a business decision about who benefits most from access and who has the appropriate context to use it responsibly.</p><h4><strong><em>Permission-Aware Responses</em></strong></h4><p>Here’s where it gets important: when one asks about funds, the system queries the database using your permissions. A tax specialist sees the funds they manage. A general partner sees their own investments. The AI doesn’t have special access — it works within exactly the same boundaries as the traditional interface.</p><p>This means you can trust the assistant to show you what you’re allowed to see and nothing more. There’s no risk of accidentally exposing one client’s data to another, because the permission filtering happens at the database level, not in the AI layer.</p><h3><strong>How It Works (basically MCP server/client)</strong></h3><p>For those curious about the technology without diving into code, here’s the simplified flow:</p><p>When you type a question, it travels to our backend server, which acts as the coordinator. The backend sends your question to an AI language model (Bedrock/GPT-5/choice) along with information about what tools are available — things like “list funds,” “get balance sheet,” “search by name.”, etc.</p><p>The AI analyzes your question and decides which tools to use. If you ask about a specific fund’s P&amp;L, it might first search for the fund by name, then retrieve the profit and loss statement.</p><p>The MCP sits between the AI and your database, accessible only to authenticated, authorized requests from the backend.</p><p>When the AI decides it needs fund data, the backend makes a request to this MCP server, passing along your user identity. The MCP server then collects the data with the permissions applied and returns the filtered results.</p><p>The AI takes those results and formulates a human-readable response, which travels back through the backend to your chat window.</p><p>This separation — AI reasoning in one place, data access in another — gives us clear boundaries. The AI never touches the database — not even for a millisecond. Every data request goes through our permission-aware tool layer. This architecture makes security auditable and maintainable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t5T1CCQ-7rjtslsA2AnLrQ.png" /><figcaption>MCP isn&#39;t all that complex</figcaption></figure><p>Signing off,<br>Sandeep.</p><p>*it is anonymised data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8f644336ab1f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kotaicode/conversational-finance-ai-assistant-that-talks-to-your-fund-data-8f644336ab1f">Conversational Finance: AI Assistant that talks to your Fund Data</a> was originally published in <a href="https://medium.com/kotaicode">kotaicode</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Next.JS isn't all that bad.]]></title>
            <link>https://medium.com/kotaicode/nextjs-isnt-all-that-bad-082e7e36c81f?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/082e7e36c81f</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[nextjs]]></category>
            <category><![CDATA[front-end-development]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Tue, 11 Nov 2025 14:50:49 GMT</pubDate>
            <atom:updated>2025-11-11T15:01:38.320Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*e7sn1fKu55nKFvM3.png" /></figure><p>Though I am primarily a backend/system developer, I have worked with frontend for a long time. Coffeescript, React JS/Native when it was just in beta, Rails, plain HTML/CSS, React when it matured a bit more, etc.</p><p>Many a time, my work there had been shooting in the dark, because one (or I) couldn’t easily find the rhyme/reason in some of the things that frontend required.</p><p>I have always gotten things done, and to the client’s satisfaction*. But not always to my satisfaction.</p><p>And all that started to change with NextJS and the new APIs of react. It is very well structured that my backend mind can also grasp how things work, and program it well (and fast) to my and the client’s satisfaction.</p><p>Many speak against NextJS (and I am aware of the drawbacks). But we have a pretty large, well functioning frontend for <a href="https://www.talix.intellifunds.de/">Talix</a>; all built in NextJS.</p><p>We haven’t had any mentionable deployment issues that blocked us. We just have a simple docker based installation and that has worked for the last 2+ years without any big glitch.</p><p>Since we (or I) are (am) heavy supporters of typescript, surely all the NextJS projects are in Typescript too.</p><p>An additional gift from the Gods is tailwind. CSS had been another painful thing I hated. But with Tailwind, I “understand” even that better.</p><p>There are a ton of resources online, but what got me started was/were some of the courses from &lt;<a href="https://www.udemy.com/user/robin-lebhar/">https://www.udemy.com/user/robin-lebhar/</a>&gt;</p><p>And now, finally with AI assistants in the IDE, I am possibly 20 times more productive than how I was 8 years ago; and I thank NextJS/Typescript/Tailwind/React for that growth.</p><h4>Footnotes</h4><p>*With one exception where we had a combination of a Hugo website with login where PHP had to be used to connect to their internal systems, and what not. That one was weird.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=082e7e36c81f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kotaicode/nextjs-isnt-all-that-bad-082e7e36c81f">Next.JS isn&#39;t all that bad.</a> was originally published in <a href="https://medium.com/kotaicode">kotaicode</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Resume Tailor — from side project to Product]]></title>
            <link>https://medium.com/kotaicode/resume-tailor-from-side-project-to-product-2866db73d022?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2866db73d022</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product]]></category>
            <category><![CDATA[resume]]></category>
            <category><![CDATA[saas]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Fri, 17 Oct 2025 16:21:23 GMT</pubDate>
            <atom:updated>2025-10-20T06:15:56.699Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*omWCVLaIzZoOUhHBBH52SA.png" /><figcaption>a painkiller for the talent/staff/resource manager/s</figcaption></figure><h3>Resume Tailor — from side project to Product</h3><p>Running a software agency has many advantages — it’s overall wonderful, but not without a few pain-points.</p><p>One of the most tedious tasks is tailoring CVs for different openings. My business partner and I have always found it mind-numbing and surprisingly hard to streamline.</p><p>As one knows, creating beautiful documents with Microsoft Word can drive one mad. One wrong move and your document will explode sending shrapnel straight to your face.</p><p>And not to mention all the different versions (specialist in RESTApi, or specialist in AWS) based on different openings.</p><p>So, we just built a tool. A homegrown one that allows us to</p><ol><li>Keep track of different personas<br>2. Keep track of different projects that the team and individual members have done (a very detailed portfolio).<br>3. And a Resume Builder with AI</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aTYk1zwW6_mpg2DpBN57Dg.png" /><figcaption>directly give the measurements to the tailor</figcaption></figure><p>All one has to do is log in, choose the team member who’s the best fit, and create a resume — then either paste the URL to the requirements (or paste the requirements text), and ask the system to create it.</p><p>(One can even specify how detailed things should be. Just a one pager?, or a beefed up version?)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XgnTZI3oozolQgghJbUSsQ.png" /><figcaption>fine tune it the way you want it to look</figcaption></figure><p>You get a preview that you can edit live, and then export to a PDF that goes to the headhunter.</p><p>(It even adds a small self-promotion URL that will show the preview of the CV directly in the application)</p><p>Like many side projects, it began as a small fix for an internal pain — and it’s quickly turning into something much bigger — blossoming to a full-fledged product.</p><p>You can read more about Team-Tailor here: <a href="https://teamtailor.kotaicode.solutions/what-is-team-tailor">What is Team-Tailor</a>.</p><p>Or check out our pitch deck here: <a href="https://teamtailor.kotaicode.solutions/pitch-deck">https://teamtailor.kotaicode.solutions/pitch-deck</a></p><p>Signing off,<br>Sandeep</p><p>(This article is follow up to <a href="https://medium.com/kotaicode/why-most-side-projects-fail-and-how-to-build-one-like-a-real-product-e8d8338fe2d4">Maryam’s post</a>.)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2866db73d022" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kotaicode/resume-tailor-from-side-project-to-product-2866db73d022">Resume Tailor — from side project to Product</a> was originally published in <a href="https://medium.com/kotaicode">kotaicode</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Crossplane & Composition: Taming Secrets at Scale]]></title>
            <link>https://medium.com/kotaicode/crossplane-composition-taming-secrets-at-scale-5df34bb92d83?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/5df34bb92d83</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[crossplane]]></category>
            <category><![CDATA[composition]]></category>
            <category><![CDATA[clients]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Thu, 18 Sep 2025 15:26:41 GMT</pubDate>
            <atom:updated>2025-09-19T09:35:58.824Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Arh9KKmUDn08XWoK6AM09Q.png" /></figure><p>In one of our client engagements, the development team found themselves in a bind. Running Kubernetes on AWS, they had to juggle <strong>dozens of apps</strong> — -each needing its own set of secrets, each demanding fresh databases on demand, and all under the watchful eyes of policy restrictions.</p><p>Keeping secrets in Git was a non-starter: a recipe for disaster.</p><p>Meanwhile, spinning up new databases meant repeatedly looping through AWS consoles or APIs, creating credentials, and somehow piping them back to the cluster. Editing secrets was another battle — -devs lacked direct cluster access, and ops didn’t want to hand out broad AWS permissions.</p><h3>Enter Crossplane</h3><p>Crossplane, in essence, is <strong>Lego for Kubernetes</strong>. Install it once, bring in the providers you need, and suddenly your cluster speaks the language of AWS, GCP, Azure — -whatever cloud you work with. Each provider extends Kubernetes with new CRDs (custom resource definitions). Want an RDS instance? Declare it. Need an S3 bucket? Apply a manifest.</p><p>But raw building blocks are not enough. You don’t want every developer hand-crafting YAML for RDS and IAM policies. That’s where <strong>Compositions</strong> come in.</p><h3>Composition: Blueprints for Lego</h3><p>A composition is a blueprint that tells Crossplane what “one thing” really means. You define it once — -either as YAML or in Golang for more flexibility — -and every time a dev asks for that “thing,” Crossplane reliably assembles the right resources under the hood.</p><p>It’s the difference between asking for a <strong>car</strong> versus assembling all the bolts, axles, and gears yourself.</p><h3>Our Solution</h3><p>Armed with these tools, we built two key compositions for our client:</p><ol><li><strong>Secrets Management</strong></li><li><strong>Database Provisioning</strong></li></ol><h3>Secrets Made Simple</h3><p>The “secrets” composition allowed developers to declare a single YAML file:</p><pre>apiVersion: secrets.example/v1<br>kind: SecretBundle<br>metadata:<br>  name: payment-service-secrets<br>spec:<br>  tag: team-payments</pre><p>That’s it.</p><p>Crossplane took it from there:</p><ul><li>Provisioned an AWS Secrets Manager vault.</li><li>Linked it to the cluster via External Secrets Operator.</li><li>Enforced access policies tied to the developer’s custom tag.</li></ul><p>Now, developers with limited AWS permissions (the permissions were tied, restricted to the tags of their team, and the ones created in AWS) could create, update, or delete secrets safely — -without ever touching the cluster directly. The secrets flowed automatically into Kubernetes, ready to be mounted in deployments and pods.</p><h3>Databases On Demand</h3><p>The database composition followed a similar path: developers requested a DatabaseBundle resource, and Crossplane spun up RDS instances with associated secrets, wired directly back into the cluster. No waiting for ops, no hunting through AWS consoles.</p><h3>Why This Worked</h3><p>This approach solved three pain points at once:</p><ul><li><strong>Security</strong>: Secrets stayed out of Git.</li><li><strong>Self-service</strong>: Developers had the power to provision without full AWS access.</li><li><strong>Scalability</strong>: Adding more apps was as simple as applying YAML.</li></ul><p>And because compositions are just blueprints, the same logic can generalize to other providers and even other clouds. Want to try it on GCP? Swap the provider. On Azure? Same story.</p><h3>In case you wonder…</h3><ul><li><strong>Scaling &amp; Namespaces</strong>: Be deliberate with naming. Within a namespace, you need conventions to avoid collisions.</li><li><strong>Crossplace/Composition</strong>: There is a small learning curve. Once mastered, it’s surprisingly elegant.</li><li><strong>AWS Policies</strong>: Spend the time upfront — -fine-grained IAM is critical to keep the model safe.</li><li><strong>Secret Rotation</strong>: Currently manual, but ripe for extension using AWS’s native rotation features.</li><li><strong>Language Choice</strong>: YAML is fine to start, Golang unlocks more customisation.</li><li><strong>Multi-cloud</strong>: This pattern is portable. The trick is picking the right providers.</li></ul><h3>Wrapping Up</h3><p>Is this the <em>only</em> way to solve secrets and database provisioning at scale? No. But it’s one of the <strong>coolest</strong> ways we’ve found — -balancing security, developer autonomy, and operational sanity.</p><p>If tomorrow you wake up with a similar headache in your dev department, you’ll want a team that can wield Crossplane and Compositions with confidence.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5df34bb92d83" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kotaicode/crossplane-composition-taming-secrets-at-scale-5df34bb92d83">Crossplane &amp; Composition: Taming Secrets at Scale</a> was originally published in <a href="https://medium.com/kotaicode">kotaicode</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP for Enterprise: Why Context-Driven AI Matters for Your Organization]]></title>
            <link>https://medium.com/kotaicode/mcp-for-enterprise-why-context-driven-ai-matters-for-your-organization-354dfb2a2c41?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/354dfb2a2c41</guid>
            <category><![CDATA[enterprise]]></category>
            <category><![CDATA[mcp-protocol]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Mon, 25 Aug 2025 15:13:34 GMT</pubDate>
            <atom:updated>2025-08-25T15:13:34.105Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E5Q78v0P4i6ggz1IbZU-sA.jpeg" /><figcaption>Context is king</figcaption></figure><p>Many organisations that tried AI in workplace have had it hard. Why? Lack of enterprise integration.</p><p>Your team’s AI Assistant should be able to answer questions about your business and get a useful, secure answer — <strong>relevant to your business</strong>.</p><p>That’s the promise of context-driven AI. Model Context Protocol (MCP) is a new (10 months old) approach making this possible by giving AI just the right information. An AI that doesn’t know everything can be a good thing!</p><p>I’ll try to answer why a generic AI alone isn’t enough, and how controlling context improves security, focus, and even cost.</p><h3>What is MCP in Plain Language?</h3><p>Model Context Protocol (MCP) is an open standard that securely connects AI systems to enterprise data and tools, so the AI can work with the right context at the right time.</p><p>It will allow you to create tailor made AI agents that will suit your need — to have account privacy, own models, own model hosting, with good performance, along with content control and governance</p><p>With it, you can connect your AI systems with your company’s data and tools — like a universal adapter to plug into various databases, apps, and services *in a controlled way*.</p><p>In upcoming articles, we’ll dive into hands-on developer guides and executive-level governance practices.</p><h3>Why Just an LLM Isn’t Enough</h3><p>LLMs are powerful, but generic ones don’t know your proprietary data (and they shouldn’t either). Ask about an internal project, and you’ll get either “I don’t know” or a confident but wrong guess.</p><p>Employees often paste documents into the AI — slow, risky, and still manual. Without system integration and integrated workflow, an LLM can’t update records or pull real numbers; it’s just a Q&amp;A tool with gaps.</p><h4>Benefits of having specialised agents:</h4><p>• Security: You control exactly what the AI can access — nothing is exposed by default. Permissions and audit logs keep sensitive data safe.<br>• Focus: The AI sees only relevant, current information, avoiding distractions and hallucinations, leading to sharper answers.<br>• Cost Efficiency: Feeding only necessary data reduces compute usage and costs. MCP keeps queries lean without sacrificing quality.</p><h3>Your AI Doesn’t Know Everything — and That’s a Good Thing</h3><p>It may sound counterintuitive, but in enterprise AI, less is more. By design, the AI only sees the data you provide for each task. This limits risk: if the AI errs or is compromised, only that slice of data is affected — not your entire system. It also avoids leaking confidential info in the wrong context. With MCP, you gain AI’s power without losing control: the assistant stays within the boundaries you set. Its lack of full-system access is a feature, not a bug — it ensures oversight and makes AI safe, reliable, and useful today.</p><p>Would you give any employee full access to all your systems?</p><h3>Quick Example: AI in Finance Fetches Report Data</h3><p>A concrete example: imagine a fund Accounting/Taxation software where each GP can chat with the platform for details about their own funds, and the system will base the answer purely on their data. The MCP server/client mechanism makes sure that the data read from the DB, or sent to the (self hosted) LLM will be just the relevant/GP&#39;s data.</p><p>Or a generic example — a manager asks, “Which regions performed best last quarter?” Without context, an AI would guess. With MCP, it securely connects to the analytics database, pulls live sales data, and replies in plain English: “Northeast and Southwest led, 15% above target; others were on track.” The manager gets instant, accurate insight without dashboards or delays. MCP ensures the AI only accesses approved data — showing how context-driven AI saves time and sharpens decisions.</p><h3>Beyond Finance: Other Industries Benefit</h3><p>Context-driven AI benefits any data-sensitive field. Finance, healthcare, legal, retail, and support can all use these agents to give AI access only to the right slice of information — patient records, case files, purchase history — while keeping everything else secure. The result: accurate, compliant, and up-to-date answers without risking exposure.</p><h3>Stay Tuned for What’s Next</h3><p>Context-driven AI with MCP will reshape work — securely and efficiently. We’ve covered (parts of) the “what” and “why”; next comes the “how.” In upcoming articles, we’ll dive into details — both for developers and decision-makers — showing how to build and govern intelligent assistants and workflows in your enterprise.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=354dfb2a2c41" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kotaicode/mcp-for-enterprise-why-context-driven-ai-matters-for-your-organization-354dfb2a2c41">MCP for Enterprise: Why Context-Driven AI Matters for Your Organization</a> was originally published in <a href="https://medium.com/kotaicode">kotaicode</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ten Commandments for working from home]]></title>
            <link>https://medium.com/@sillyfellow/ten-commandments-for-working-from-home-37bbdf8e9316?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/37bbdf8e9316</guid>
            <category><![CDATA[work-from-home]]></category>
            <category><![CDATA[wfh]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Tue, 17 Mar 2020 08:17:43 GMT</pubDate>
            <atom:updated>2020-03-17T08:25:52.828Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bCFC1SMRa5LapgaeUixo6A.jpeg" /></figure><ol><li>Start early — at the time you usually leave to work. Don’t squander the travel time.</li><li>Have a dedicated room, or at least a table for your work (keep charging cable ready here)</li><li>Try to keep your calls to one session — before or after noon. Wear pants at least during this session.</li><li>Use the saved travel time for lunch. A relaxed lunch; a ted talk, or a twenty minute episode of something. Going for a walk in times without COVID-19.</li><li>Call it a day when you usually finish working.</li><li>Travel time is again at your disposal. Take at least half of it to sharpen the saw — improve your skills.</li><li>Cook in the evening. Tomorrow’s lunch won’t prepare itself.</li><li>If your partner works from home too, align your calendars on Sunday evening, to avoid conflicts in the study.</li><li>If childcare is necessary, roughly divide the hours the previous day itself.</li><li>DO NOT FORGET to show your boss that you’re actually more productive, and the time you put into improving skills will pay off in future WFH — even when Corona is a distant memory.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37bbdf8e9316" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SSH to WSL from terminal emulator]]></title>
            <link>https://medium.com/@sillyfellow/ssh-to-wsl-from-terminal-emulator-3de68015dcea?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/3de68015dcea</guid>
            <category><![CDATA[ubuntu]]></category>
            <category><![CDATA[ssh]]></category>
            <category><![CDATA[wsl]]></category>
            <category><![CDATA[wsl-2]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Tue, 04 Feb 2020 19:05:05 GMT</pubDate>
            <atom:updated>2020-02-04T19:38:56.393Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/883/1*r7SN-U1ah8YKwZvEyJuEuA.png" /></figure><p>So, you have setup WSL ( Windows Subsystem for Linux) and are not happy with the “built in”emulator. In my case, I wanted to use an emulator and ssh to the Ubuntu in wsl2.</p><p>I hit a few road blocks, but here’s how one could get around them.</p><h4>Install ssh</h4><pre>sudo apt install openssl openssh-server</pre><p>Windows will ask you whether you want to allow ssh through the firewall. Say yes.</p><h4>Fix the configuration</h4><p>Open the configuration file at /etc/ssh/sshd_config and do the following.</p><ol><li>Change the port from 22 to 2222. (or another one of your choice)</li><li>(optional) Make key authentication to no and password auth to yes</li></ol><pre>PubkeyAuthentication no<br>PasswordAuthentication yes</pre><p>Only if you haven’t setup the keys</p><h4>Start the server</h4><pre>sudo service ssh --full-restart</pre><p>You might see that the (ssh) keys are missing, like below</p><pre>Could not load host key: /etc/ssh/ssh_host_rsa_key<br>Could not load host key: /etc/ssh/ssh_host_ecdsa_key<br>Could not load host key: /etc/ssh/ssh_host_ed25519_key</pre><p>If so, create the keys like below</p><h4>Create the keys</h4><pre>ssh-keygen -A</pre><p>And restart the server once more (see above)</p><h4>Good to go</h4><p>Go to the terminal of your choice, and ssh to localhost.</p><pre>ssh -p 2222 user@localhost</pre><p>It’ll ask for the password, and might be a bit slow to let you in.</p><p>Be patient, you will be rewarded.</p><p>Cheers, Sandeep.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3de68015dcea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Merge Channels in Go(lang)]]></title>
            <link>https://medium.com/@sillyfellow/merge-channels-in-go-lang-cfef9d76b48b?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cfef9d76b48b</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[go]]></category>
            <category><![CDATA[channel]]></category>
            <category><![CDATA[golang]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Mon, 08 Jul 2019 08:15:11 GMT</pubDate>
            <atom:updated>2019-07-08T08:50:21.650Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently, I had to deal with receiving several events from Kafka. There were about 6 different types of events, but internally, the ways I had to deal with them were quite similar — so I decided for a generic method.</p><p>From Kafka, I was receiving events in individual channels (one channel per event type)</p><p>As my handling of the events was the one generic method, I wanted to take all the channels, merge them into a single channel which has all the events from all channels, combined*, from which I can pick one at at time**, and deal with it.</p><p>Doing it in Go was a breeze. Seriously, in 10 lines, all was done.</p><p>Here’s how I did it. (well, business events are replaced by integers here ;) )</p><p>I have two channels — one which delivers me odd numbers, and another one which delivers even numbers. And here’s how one can merge them.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/7e24477574bff7bd8a7e2c06f8686d25/href">https://medium.com/media/7e24477574bff7bd8a7e2c06f8686d25/href</a></iframe><p>Note that the function accepts a slice of channels — so, it doesn’t matter whether it’s just two channels, or 100.</p><p>For completion, here’s my sample program which contains the channel creations as well.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/c4a5ecaf93c3243049cbb9248cc3997a/href">https://medium.com/media/c4a5ecaf93c3243049cbb9248cc3997a/href</a></iframe><p>Run it as follows:</p><blockquote>go run channelmerge.go</blockquote><p>And you’ll see all whole numbers being printed on your screen. (warning: infinite loop)</p><p>If you look closely, you’ll see that they are not all in order — of course, the channels are not synchronised with each other.</p><p>Cool. ah?</p><p>Signing off,<br>Sands.</p><blockquote>*Combining the channels would give me a better solution than having 6 different loops to do the same thing, and having to add more such loops when I have more events-types</blockquote><blockquote>**Naturally, at some point of time, I would want to speed that up,.. That we can do in another post for ThreadPooling</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cfef9d76b48b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rustup Frankfurt (2019–04–25)]]></title>
            <link>https://medium.com/@sillyfellow/rustup-frankfurt-2019-04-25-e5ec279779ba?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5ec279779ba</guid>
            <category><![CDATA[rust]]></category>
            <category><![CDATA[meetup]]></category>
            <category><![CDATA[frankfurt]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Mon, 29 Apr 2019 09:23:21 GMT</pubDate>
            <atom:updated>2019-04-29T09:23:21.011Z</atom:updated>
            <content:encoded><![CDATA[<p>The <a href="https://www.meetup.com/Rust-Frankfurt/events/260244823/">first meet</a> of the Rust Frankfurt happened quite successfully.</p><p>The enthusiasm shown by the members during the meetup was an actual reflection of the same which we saw earlier in the meetup group.</p><p>Though new to Rust, all the 8 attendees were seasoned programmers who have deeper understanding of concepts, which led to small interesting discussions and this sharing simply elevated the quality of the event.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AX1Y8ckXHSngY87FhUddlA.png" /><figcaption>happy rustaceans of Frankfurt</figcaption></figure><h4>How it happened?</h4><p>Around 19:00, the Rustaceans started flocking around at <a href="http://kotaico.de/">Kotaicode</a> offices at Bornheim.</p><p>After the first introductory round and coffee, we moved on to setting up Rust on our laptops (A repository with instructions to have quick setup or Rust — even dockerised setup, should someone choose to — was readied for the meetup : <a href="https://github.com/kotaicode/rust-frankfurt-meetup">https://github.com/kotaicode/rust-frankfurt-meetup</a>).</p><p>Then we made sure to do our customary Hello World. Of course, our world view is limited to the Rustaceans of Frankfurt. :)</p><p>Then we successfully went through the first two chapters of the <a href="https://doc.rust-lang.org/book/index.html">Rust-book</a>.</p><p>By the end of the session, we all had a better understanding of the lay of the Rust land, than before.</p><h4>What next?</h4><p>We decided to have the meetups in a biweekly frequency, on Wednesdays.</p><p>And instead of going through the tutorials, we would be going to some sample code/project prepared by the members. The <a href="https://www.meetup.com/Rust-Frankfurt/events/261010432">next event</a> is already announced.</p><p>Looking forward to seeing you there.</p><p>Signing off,<br>Sands.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5ec279779ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[(Re)ranking is amortized O(1); and atomic]]></title>
            <link>https://medium.com/@sillyfellow/re-ranking-is-amortized-o-1-and-atomic-fbb182354d29?source=rss-557be62cc30b------2</link>
            <guid isPermaLink="false">https://medium.com/p/fbb182354d29</guid>
            <category><![CDATA[rankings]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[ordering]]></category>
            <dc:creator><![CDATA[Sillyfellow]]></dc:creator>
            <pubDate>Sun, 22 Jul 2018 18:43:13 GMT</pubDate>
            <atom:updated>2018-07-22T18:43:13.291Z</atom:updated>
            <content:encoded><![CDATA[<p>When you build APIs, sometimes you need to provide a re-ordering API.</p><p>Perhaps a new item is added to the collection — which always has to be at one end. Or there is a cool drag and drop re-ordering in the front-end which you need to support via re-ordering api.</p><p>Obviously, if a new item is to be added at the bottom of the list, it is easy<br> — set the rank of the new one to be the highest rank + one.</p><p>No other ranking is obvious — imagine moving the item from position <strong><em>k</em></strong> to the<br>first — one would end up changing the rank of every item between <strong><em>0</em></strong> and <strong><em>k</em></strong>.<br>(See <a href="https://softwareengineering.stackexchange.com/a/331650/148333">this link</a> for a stack exchange comment suggesting this)</p><blockquote>(Re)ranking one is re-ranking many‽</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cr3QZjVmYP-TZ-qZ_spMAw.png" /><figcaption>Wherever a new item comes, all the ones below it until the gap formed, or till the end must be re-ranked. It’s O(n) my friend.</figcaption></figure><p>This has two problems to start with:</p><blockquote>1. It is <strong><em>O(n)</em></strong> per re-ranking</blockquote><blockquote>2. It is <strong><em>not atomic</em></strong> — if the re-ranking fails at any point, the resulting order may not be correct anymore.</blockquote><h4>It’s all relative</h4><p>If we see ranking as relative positioning instead of an ordered sequence of<br>numbers, we can do with just one operation.</p><p>I.e., if an item has to be placed at position 2, all we need, and all we do is just that.</p><p>Imagine, we have 10 items. We give them <strong><em>ranks</em></strong> 10, 20, 30, …, 100 to start with. When the item from, say, position 7 wants to move to position 2, we will change its <strong><em>rank</em></strong> to 15.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-RJd_esUG7t1wej-IcwrdA.png" /><figcaption>Moving an item is just a single operation, finding the mean of the ranks of the new surrounding items and assigning it.</figcaption></figure><p>Done. That’s it.</p><p>The new <strong><em>ranks</em></strong> are 10, 20, 30, …60, 15, 80, 90, 100 — which, if you sort by rank, will give the desired order.</p><p>It will work well for 2–3 times, but then it will break down.</p><p>If the 8th one should move up to position 2, we’ll make its <strong><em>rank</em></strong> to 12. <br>The next one to go to position 2 will get the <strong><em>rank</em></strong> 11. <br>For the next one we are blocked.</p><h4>Formally</h4><p>It can be stated that, if items are always moved to one specific position, we will have the first collision after <strong><em>log</em></strong>₂<strong><em>(m)</em></strong> times where <strong><em>m</em></strong> is the size of the window between two consecutive items.</p><p>So, a larger window will help reduce clashes.</p><p>This problem is mitigated by two means.</p><p>1. Clever windows/positioning initially<br>2. Re-spreading when a clash happens</p><h4>Clever Positioning</h4><p>We humans might be a little limited when it comes to numbers, but the computers can easily go till <strong><em>2⁶⁴</em></strong>. If we cleverly distribute the rankings between <strong><em>0</em></strong> and <strong><em>2⁶⁴</em></strong>, it will take a LONG time indeed before there is a clash of any two items.</p><p>If we have the rough number of expected items in the list, it would help nicely. The placing of consecutive items can be done perfectly with a difference of <strong><em>whole-window-size / (2 * expected-count)</em></strong>.</p><p>Imagine we expect about <strong><em>1000</em></strong> items in the list.</p><p>For the first item, we give the <strong><em>rank</em></strong> to be the mid-way of our expected window — i.e., <strong><em>2⁶⁴ / 2 = 2⁶³</em></strong>.</p><blockquote>NOTE: Choosing the middle will let us grow to either side of the list easily.</blockquote><p>For each next item, we place it at a distance of <strong><em>2⁵²</em></strong> away from the closest<br>one. This will work clearly, because we have <strong><em>2⁶³ / 2⁵² = 2¹¹</em></strong>, which is<br>larger than twice the expected need.</p><blockquote>NOTE: 2 times the expected size will give us enough space for the whole list on either side of the mid-way — in case we want to add them all on one side.</blockquote><p>With this, we can spread out the whole list with a window size of <strong><em>2⁵²</em></strong>. In<br>other words, it means, it will take <strong><em>52</em></strong> re-rankings to the same exact position<br>before we have any collision.</p><p>As luck would have it, such repetitive actions — where every subsequent item is moved to the SAME position — are rare, and the users would be ranking items around to various positions.</p><p>This means windows will almost never go too small to create a need for re-ranking.</p><blockquote>NOTE: If we do not have an indication of the size, go happily with 2³². :)</blockquote><h4>Ranking Algorithm</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/8394581c3ff5a47b6aca18023f762910/href">https://medium.com/media/8394581c3ff5a47b6aca18023f762910/href</a></iframe><p>Obviously, we really don’t need all the siblings in the memory. If it is in a<br>database, we can read only the few needed ones. I leave those details to you.</p><h4>Rank Spreading</h4><p>In the worst case scenario, at some point of time, if/when we have a collision, we resort to a method similar to the above and spread the items equally spaced in the whole space.</p><p>The spreading can happen from one end, and can be done per item, not all must be done in a go.</p><p>Even if the loop breaks accidentally, the order will be preserved, and one doesn’t have to worry about losing the work done — unlike the traditional ranking.</p><h4>Analysis and Practicality</h4><p>I have not done a proper <a href="https://en.wikipedia.org/wiki/Amortized_analysis">amortized analysis</a> of the operations, but I believe we don’t need a detailed analysis to see that the re-orderings will be very few and far apart, and each ranking needs only one <strong><em>write</em></strong> action — leaving the algorithm to be <strong><em>O(1)</em></strong> amortized for a single re-ranking.</p><p>The code may not be too obvious, but it is <em>perhaps</em> better testable than the traditional ranking.</p><p>— Sillyfellow</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fbb182354d29" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>