I recently started reading about MCP (Model Context Protocol) to understand how AI applications interact with external tools and data. But after a while, I realized that theoretical definitions weren't really helping me understand how it works. So I decided to build a small MCP + SQLite employee database assistant and learn the basics by actually putting the pieces together. Along the way, I learnt about MCP servers, clients, tools, tool discovery, and how AI can interact with a database through a standardized interface. One takeaway: with AI-assisted coding, writing code is becoming less of a bottleneck. Understanding the problem, designing the architecture, and making the right engineering decisions are becoming increasingly important. Small implementation, but a good introduction to a space I’m excited to explore further. 🔗 GitHub: https://lnkd.in/dmYxVE_F Opinions and feedbacks are welcome :) #MCP #AI #AgenticAI #SoftwareEngineering

Claude Code 2.1.282, released September 24, 2026, tightens a specific security boundary: what project configuration can change about telemetry. The update ignores project and local settings that enable OpenTelemetry export, redirect its endpoint, or capture content. It also adds notices identifying ignored telemetry variables. In plain language, configuration that comes with a project loses authority over these sensitive telemetry choices. The notices make it clearer when telemetry variables are being ignored. The practical benefit is a narrower, more explicit trust boundary. This is not evidence that Claude Code is universally secure or telemetry-free; it is a targeted configuration-security improvement. For a third-year computer science student, this offers a concrete way to study security in developer tools: configuration files are part of the attack surface, not just convenient preferences. A useful learning exercise would be to design a settings system that separates repository preferences, such as formatting, from user-controlled security settings. Then write tests using an untrusted sample repository: can its configuration enable telemetry export, change the destination, or turn on content capture? Does the tool explain which settings it ignored? That connects an AI coding tool update to a broader software-development lesson: settings need clear limits on their authority. Where would you draw the boundary between project preferences and user-controlled security settings in a developer tool? Source: https://lnkd.in/g5pm3cdD #DeveloperTools #SoftwareSecurity

Current assessment of the best models: For a while now I rate these models largely as good as the best grad students I studied with at Harvard participating in advanced seminars. They have superior memory and are able to bring in even more relevant material from the literature. So they are superb co-workers. One problem with them is that they are now being remarkably pro-active, which means that they try to anticipate what you need or want, and from a given project description they can considerably augment the work plan with guard rails, rules, limits, and details, going far beyond what you had in mind or even imagined. And much of this can be undesirable and create unwanted obstacles on the way to your goal. Back in my previous life as a machine learning software developer, I had a supremely talented developer who, I thought, overthought and overplanned his work. I wanted to look over something he was working on, and discovered that the files I needed to review sat in a directory structure 7 levels deep. He wanted, of course, to separate our work by target operating system (we supported Windows, Mac, Linux, proprietary UNIX), and he also anticipated a number of possible future variations, and so made provision for speculative branches of work that in fact never came to be. When I see some of the work these agents do, they remind me of that programmer. In order for me to stay in control, I have to narrow and simplify their vision. If not, we arrive at a complexity that turns out to be very difficult to navigate. In terms of preference Anthropic vs openAI I am oscillating, one week finding Anthropic superior and the next OpenAI. So I continue to use them both having them critique each other's work. This how I recently discovered that a 3 sentence clarification I issued turned into 37 new rules, guardrails, cautions, decision gates, etc. Now I have to guide the simplification.

In last week's database modelling class, I followed along on the whiteboard - but without any hands-on time in the modelling software. The program runs through the university VPN (licensed app), and a personal VPN I'd left running on my laptop got in the way. I lost the hands-on part of the session - but gained an idea, and a few days of building my own data modelling app with Claude. StrataSQL turns a description into a database design, then into real SQL - right in the browser: ① Watch - a model is built step by step, every decision explained ② Practise - build it yourself; the trainer checks your model and tells you what's missing You can also paste your own exercise or open your own model. Working with AI is both methodical and creative process: a spec, acceptance criteria, test cases, and a cross-check against my own class models. App runs in any browser, Windows or Mac, and keeps working offline once opened. If you know someone learning databases, feel free to share it. https://model.quaera.app * An unofficial student project: where it differs from the course, the course is right. #DataModeling #Databases #SQL #AI #LearningByDoing

The learner proudly said, “Form bana diya.” The screen looked good. Name field. Email field. Password field. Sign Up button. Then came the next question: “Data jayega kahan?” Silence. Another learner suggested: “Cloud mein?” Fair answer. But… where in the cloud? Which table? What fields? How will users connect to orders? How will products connect to order items? Which value becomes the primary key? Which relationship needs a foreign key? Suddenly, the form was no longer just a form. It was part of a complete data model. That is an important shift in software development. Problem: Learners often focus on visible screens without understanding how application data should be structured behind them. Solution: Learn database tables, relationships, keys, storage, and retrieval as part of the application-building process. Result: Stronger backend thinking, better application architecture, and more meaningful full-stack project experience. At Dousoft EduVerse, the goal is to help learners understand both sides of an application: What the user sees. And what the system needs to remember. Because “cloud mein daal denge” is not quite a database architecture. Database table furniture nahi hai. UI ke peeche data thinking bhi hoti hai. 🌐 https://lnkd.in/g-FYJePF 📲 WhatsApp: +91 88776 39698 📧 info@dousofteduverse.com #DousoftEduVerse #DatabaseDesign #SQL #BackendDevelopment #FullStackDevelopment #Programming #SoftwareDevelopment #DataModeling #CodingSkills #ProjectBasedLearning #StudentLearning #DeveloperSkills

Day 29 Of Learning Data Science Until I Land A Job. Most apps built on LLMs are paying for the same text over and over. Every call resends the same long instructions, and you get charged for reading them every time. In my own apps the system prompt is the biggest part of every request. Rules, examples, format, all the stuff that never changes. The user's actual message is a line or two at the end. So 90% of what I was paying for was the model reading the same instructions it read a second ago. - cache_control={"type": "ephemeral"} That one line on the fixed part of the prompt tells the API to keep it. Next call within a few minutes, it doesn't reprocess it, it picks up where it left off. Cached reads cost a fraction of the normal price and come back faster too. The catch is the order matters. The cache only works on the start of the prompt, up to the first thing that's different. So anything that changes, like the user's name or today's date, has to go at the end. Put a timestamp at the top and you've broken the cache on every single call without realising. For anything with a long fixed prompt and lots of users, it's one of the biggest cost cuts available and barely any code. Anyone using this properly, how much did it actually knock off your bill? #DataScience #AIEngineering #LLMs

A fast function can still sit inside a slow system. That’s why I see value in learning Data Structures & Algorithms and System Design side by side. DSA helps us reason about the work inside a component. System design helps us reason about how components work together. Consider a hypothetical autocomplete feature: At the DSA level: • How should we store and search prefixes? • Would a trie or a sorted array suit the workload? • How do we select the top suggestions efficiently? • What are the time and memory trade-offs? At the system-design level: • Where does the search run—on the device or a server? • How do new suggestions become available? • What should we cache, and when should it expire? • What happens when the network is slow or unavailable? These decisions influence each other. A trie might make prefix lookup efficient, but its memory footprint still matters on a phone. A cache might reduce repeated work, but it also introduces questions about freshness and invalidation. My suggested learning loop: 1. Solve a small algorithmic problem. 2. Explain its time and space complexity. 3. Place it inside a realistic application. 4. Add one constraint: more users, limited memory, changing data or network failure. 5. Revisit the original choice. For example, after learning heaps, explore a leaderboard. After learning queues, sketch a background-job system. After learning graphs, design a dependency explorer. The goal is to connect each data-structure choice to the behaviour users experience. Learn how to make one operation efficient. Then learn what happens when many people depend on it. #DataStructures #SystemDesign #SoftwareEngineering

🚀 Built My First Local AI Research Assistant! I built a private, offline research assistant inspired by NotebookLM that runs locally on my laptop. 💡 What can it do? • 📂 Upload and work with PDFs, notes, CSV, JSON, and Markdown files • 🤖 Ask questions and receive answers powered by a local LLM through Ollama • 🔍 Generate document-grounded answers with source references • 🧠 Store notebook-specific memory for project context and response preferences • 🔒 Keep documents and data stored locally, with privacy at the core After the initial setup and model download, the core functionality works without a cloud account or an active Internet connection. This gives me greater control over my data and reduces dependence on external AI services. The assistant uses Meta’s Llama 3.2 3B model through Ollama. I chose this lightweight model because it can run locally on my 16 GB RAM laptop without requiring a dedicated GPU, while still providing useful performance for document-based question answering and summarization 🔗 Explore the project on GitHub: https://lnkd.in/d-RYkvYk #LocalAI #AI #LLM #Ollama #RAG #GenerativeAI #AIProjects #BuildInPublic

Confused about which tech domain to choose? 🤔 Web Development, Machine Learning, or Data Science? Don’t pick a domain just because it’s trending. Start with your strengths, interests, and career goals. Swipe through to understand which domain might be the right fit for you. 👉 #TechCareers #CareerGuidance #DataScience #MachineLearning #WebDevelopment

Science is hard by design. Your data workflow shouldn't be. You spend weeks shaping a hypothesis. You run protocols that take days to get right. You troubleshoot at the bench until the experiment finally works. And then the weekend goes to the analysis... Exporting files. Matching sample names across spreadsheets. Copying numbers from one place to another. Rebuilding a dataset that already exists, just scattered across a notebook, a shared drive and someone's laptop. On top of that, getting a pipeline to run can mean learning Linux, bash, containers or the cloud. Useful skills, but not the reason you got into science. Your time is best spent on the questions only you can ask, not on moving data from one place to another. I help labs streamline that part: map how your data flows today, find where time gets lost, and automate the steps that don't need a scientist. What's the analysis task that eats most of your week? If you'd like to make it shorter, let's connect.