How to Build Your Own AI from Scratch

🚀 AI Learning Journey : Building My First AI Agent! This week, I continued my AI Engineer roadmap by creating my first AI Agent — a lightweight assistant capable of understanding user intent and performing real actions like recording details, handling questions, and sending notifications. 💡 Key Highlights: ✅ Built using OpenAI, Gradio, and Pushover APIs. ✅ Integrated real-time notifications with environment-based security using dotenv. ✅ Added support for reading PDFs using pypdf — setting up for document-aware agents in upcoming projects. ✅ Explored how LLMs call custom tools (functions) dynamically to perform structured tasks. 🔍 Key Learnings: ✅ Learned how to design agent tools with JSON schemas for interaction. ✅ Understood how to integrate third-party APIs and manage secrets effectively. ✅ Built a modular structure that can easily scale into a multi-agent system. This marks the start of building real-world AI agents step-by-step — next up: connecting agents for collaborative reasoning ⚙️🤖 Thanks Ed Donner for such a wonderful course on AI Agents #AI #LearningJourney #OpenAI #Python #AIEngineer #AIagents GitHub Link : https://lnkd.in/gAcNFBXc Medium link : https://lnkd.in/df_SVc-c Link to my App : https://lnkd.in/dMc7GrGP

🚀 I just built an AI app that lets you chat with your documents — in real time! Ever wished you could upload multiple PDFs, Word, or TXT files and just ask questions about them naturally? Now you can. 💬 I created a Gradio-based AI interface powered by LangChain and OpenAI (GPT-4o-mini) that: - Accepts multiple documents (PDF, DOCX, TXT) - Summarizes each one with live streaming responses - Stores their content in a vector database (Chroma) for retrieval - Keeps chat memory for smooth, context-aware conversations - Lets you reset, refresh, and re-upload knowledge dynamically What I loved most about this build was combining streaming, retrieval-augmented generation, and conversational memory into one clean experience. It taught me a lot about document chunking, embedding performance, and how to design a truly interactive AI workflow. 💡 Next step: adding support for long-term memory and multi-user sessions — so each person can have their own private chat knowledge base. 🔗 GitHub Repo: https://lnkd.in/dFN-46tJ 👉 Would you use a tool like this for research, studying, or document analysis? I’d love to hear how you’d improve or extend it! #AI #LangChain #gradio #OpenAI #MachineLearning #python

Day 1 of Kaggle x Google’s “5 Days of AI” and we are officially leveling up our agents. Today’s focus is on moving from a Level 1 agent that simply responds to prompts to a Level 4 agent that can plan and act independently. We started with Google’s Agent Kit, a toolkit that helps developers connect AI reasoning to real-world actions like creating documents, sending messages, or running queries with just a few lines of Python. I really like how Kaggle structured the journey: • Level 1: Basic LLM that answers a single prompt • Level 2: Gains tool and API access for memory and reasoning • Level 3: Manages multi-step reasoning and task planning • Level 4: Becomes autonomous and can complete complex goals with context It is a reminder that the next wave of AI is not about what models say, but what they can do. Excited to keep building along and see how these layers evolve into a fully capable agent. Shoutout to Anant Nawalgaria and Michael Gerstenhaber for making this series both accessible and powerful. Anyone else following along? What was your favorite insight from Day 1? #Kaggle #AI #OpenAI #Google #AgentKit #PromptEngineering #MachineLearning

🎗️ Day 2 of My Google AI Agents Journey Started the day super excited — and ended it with a head full of questions (in the best way possible!). 🔧 Function Tools – Converted Python functions into agent tools 🤖 Agent Tools – Created specialist agents and used them as tools 🧰 Complete Toolkit – Explored all ADK tool types and when to use them And then came the big takeaway of the day… 💭 Still, LLMs are not perfect in Math. Not the basic arithmetic part, but anything that needs multi-step reasoning or symbolic understanding. It’s fascinating, isn’t it? These models can write poetry, generate code, and summarize research papers, yet ask them to solve a slightly twisted algebra problem, and they’ll confidently give the wrong answer. That’s the beauty and the challenge of AI — it’s still learning. While LLMs “sound smart,” they don’t think the way we do — not yet. Mathematical reasoning isn’t just pattern recognition; it’s about structure, logic, and precision. ✨ I’m genuinely excited to see how LLMs evolve to truly validate math in the future. Because mathematics isn’t just about numbers — in industry, it’s the invisible engine that keeps everything measurable, predictable, and optimizable. Super excited to catch up and reflect on my thoughts during Kanchana Patlolla’s sessions #AI #GoogleAI #LLM #ArtificialIntelligence #MachineLearning #AIAgents #MathInAI #AIJourney #DataScience #Innovation #TechLearning #AIEngineering #Kaggle

We spend hours every week on tasks that don’t need our brains, they just need better systems. Over the last month, I’ve been building small AI automations to take those off my plate: 1) Auto-summarizing meetings and Slack updates 2) Tracking metrics in real time 3) Notifying me when key numbers change 4) Generating quick insights from scattered data sources What started as a side experiment turned into a workflow that saves 5–7 hours every week. No fancy setup, just n8n, OpenAI GPT 5, and a few smart Python scripts stitched together. The goal wasn’t to “replace work.” It was to make space for thinking, for creativity, for solving harder problems. Here’s the truth: automation isn’t about scale. It’s about clarity. The clearer your process, the easier it is for AI to amplify it. #AI #Automation #DataAnalytics #WorkSmarter #Python #n8n #OpenAI #LLM #Productivity #Innovation

Been working on this for a bit – happy to finally share! ✨ Hey folks, Just wanted to share this little project I put together – it tries to guess numbers you draw on the screen. ✍️ Pretty simple idea but was fun to build! It uses a Neural Network model I trained, and honestly, it's not perfect yet. Which is kind the point! I added these little 👍/👎 buttons so you can tell it if it got the number right. If it's wrong, you can even type in the actual number (or just say "NaN" if you drew, like, a cat 🐈⬛). It saves the drawing and your correction, and the plan is to use that saved data later to hopefully teach the model to be a bit less confused. 🧠 Definitely stretched my brain getting the React canvas to play nice with the Flask/TensorFlow backend and then wrestling it onto Hugging Face Spaces. Anyway, if you're bored and feel like scribbling some numbers, give it a try! Let me know if it guesses yours right (or wrong!). 🔗 Try it here: https://lnkd.in/eiHf2aNT Always happy to hear what people think or if something breaks! Cheers. 😊 #MachineLearning #ReactJS #Flask #TensorFlow #AI #WebDevelopment #CodingFun #HuggingFace

Most AI demos just pretend to sound human. This one actually calls people. Two IT students, Sara Cubillos from Victoria University and Van Kiet Luong from Western Sydney University, decided to build an AI voice agent that can pick up the phone, talk naturally, and understand why a student didn’t show up to class. No pre-recorded audio. No fake scripts. Just real-time conversations using OpenAI + Twilio. They didn’t stop there. The AI logs everything in Google Sheets, classifies whether the student is confirmed or absent, and automatically reminds them of the course value. But the real lesson here? They faced AI latency issues, API errors, budget limits, and sleepless debugging nights. Still, they pulled it off. Because building something that works in the real world > building something that just looks cool on LinkedIn. Watch their full journey + live call demo on YouTube: https://lnkd.in/gGcFWDUt 🔗 Connect with them (they’re hireable): Sara Rojas – https://bit.ly/43mnExh Van "Kevin" Kiet Luong  – https://bit.ly/3WQSBWE #AI #OpenAI #Twilio #Python #FastAPI #Developers #VoiceAI #PortfolioProject #Innovation #TechStudents #AITools

Okay so this is actually wild—just saw that DeepSeek dropped a new open-source AI model, and it’s giving OpenAI-level reasoning on some of the toughest math and logic benchmarks. Not bad for something anyone can tinker with. #LearnWithJimil 🤔 So what’s the deal? DeepSeek-R1-Lite-Preview isn’t just another chatbot or text generator. It’s built to handle gnarly math and logic problems, like the stuff you see in AIME competitions or advanced MATH tests. That’s a big leap for open models, which usually lag behind the fancy closed ones. Imagine you’ve got a super-smart math friend who never gets tired. You throw them a problem—like a weird word puzzle, or a math olympiad question—and instead of freezing up or guessing, they actually break it down step by step. That’s the vibe here. For tech folks: DeepSeek’s new model reportedly matches OpenAI’s o1-preview on those reasoning benchmarks, which is pretty impressive given it’s open and you can play with it yourself. It’s not just fast—it’s clever. Why does this matter? Well, this could make advanced problem-solving way more accessible. Think free homework help, smarter coding assistants, or even research bots that can handle actual scientific reasoning. And because it’s open, smaller teams and indie devs get a real shot at building cool stuff. Honestly, I’m pumped but also a little skeptical. Open models catching up this fast? We’ll see how it does in the wild, but even the idea makes me want to try it out. If you’re into AI, this feels like a moment. Would you trust an open-source AI to solve your toughest math problem? Or do you still reach for the big names? Source: https://lnkd.in/gv-jEGJZ #AI #TechNews #DeepSeek #OpenSource #LearnWithJimil #MachineLearning

📅 Day 11/30 of my #GenAIDeveloperJourney — Building My First Multi-Step AI Agent (Kaggle Learnings) 🧠Today's lesson in the Kaggle AI Agents course helped me finally understand how AI Agents break down complex tasks into multiple steps — and it completely changed the way I look at agent design. ▶️ Most people think agents are just "LLMs with tools"…. ▶️ But the real magic is in their ability to plan → execute → revise → complete -without human intervention. 🕴That is exactly what businesses want: Developers who can build autonomous systems, not just prompt models. 🔍What I Learned Today: The most important thing gleaned from today's module has been the idea of Agent Planning, where an agent transforms one user request into a series of smaller, executable actions. This makes agents reliable, explainable, and capable of solving real-world tasks step-by-step instead of in one-shot guessing. Planning is the difference between: ❌ "Give me an answer" ✔️ "Let me think → break the task → solve each step → combine → answer reliably" Core Concepts I Learned from Kaggle Agents Day 3 1️⃣ Task Decomposition Agents use chain-of-thought style reasoning to convert a large instruction into micro-actions. Analyze this dataset and summarize insight: 👉Load data 👉Validating data 👉Process 👉Analyze 👉Summary Remove 👉Return result This forms the very basis of strong AI workflows. 2️⃣ Execution Traces (Agent Logs) Agents maintain a trace containing: 👉Thought process 👉Actions taken 👉Tools utilized 👉Observations received This is important for: ✔ Debugging ✔ Auditability ✔ Enhanced reasoning 3️⃣ Retries & Self-Correction Should an action fail or return invalid output, the agent: Analyzes the failure Updates reasoning Tries a different approach This is what makes agents dependable, especially in automation pipelines. #GenAI #agents #kaggle #developerjourney #openai #machinelearning #python #GenAIDeveloperJourney #Day9 #AIAgents #AgentArchitecture #LangChain #AutoGen #CrewAI #KaggleLearning #LLM #AICommunity #AutonomousAI

🌍 Agents That Learn — Not Just Prompt We’re at an interesting stage in AI. Most “agents” today don’t actually learn they just retry with better prompts. You tweak, test, and hope for smarter behavior. But what if agents could improve themselves? That’s where Microsoft’s Agent Lightning caught my attention. It’s an open-source framework that finally brings a training loop to AI agents allowing them to learn from their own interactions. Instead of rewriting your whole stack, you: ⚡ Keep your existing agent (LangChain, AutoGen, CrewAI, OpenAI SDK and even plain Python) 📈 Add a lightweight tracer 🧠 Let Agent Lightning observe, reward, and optimize over time Think of it as turning static prompt-based systems into adaptive, experience-driven agents. It’s a glimpse of where autonomous AI is headed — systems that evolve, not just respond. If you’re building in this space, it’s worth exploring 👉 https://lnkd.in/gnyQ7Jgk #AI #AgentLightning #MicrosoftResearch #AutonomousAgents #ReinforcementLearning #AIFuture #OpenSource