How we built Refactron, an AI-powered code refactoring tool

A script executes. An application decides. #ZeroToFullStackAI Day 6/135: The Principle of Control Flow. For the past five days, our code has been a simple, top-to-bottom script. It executes one line after another, no matter what. The `ValueError` from our Day 5 challenge proved this is not enough. We need a way to handle different conditions. Today, we build the "brain" of our application. This is "Control Flow". It’s the mechanism that allows our code to analyze a situation and make a decision. Our tool for this is the `if/elif/else` structure: 1. 'if' : The primary gate. It asks a `True/False` question. 2. 'elif' : The secondary gate. It *only* asks its question if the `if` was `False`. 3. 'else': The "catch-all." It runs *only* if all preceding conditions were `False`. This is the first and most fundamental tool for writing non-linear, intelligent logic. We can now create different paths for our program to follow. We've taught our code to make logical decisions. But we still haven't built the "safety net" for when it receives bad data (like the `ValueError`). That is the final piece of our foundation. Tomorrow, we build the safety net: **Error Handling**. #Python #DataScience #SoftwareEngineering #AI #Developer #Logic

A script executes. An application decides. #ZeroToFullStackAI Day 6/135: The Principle of Control Flow. For the past five days, our code has been a simple, top-to-bottom script. It executes one line after another, no matter what. The ValueError from our Day 5 challenge proved this is not enough. We need a way to handle different conditions. Today, we build the "brain" of our application. This is Control Flow. It’s the mechanism that allows our code to analyze a situation and make a decision. Our tool for this is the if/elif/else structure: if: The primary gate. It asks a True/False question. elif: The secondary gate. It only asks its question if the if was False. else: The "catch-all." It runs only if all preceding conditions were False. This is the first and most fundamental tool for writing non-linear, intelligent logic. We can now create different paths for our program to follow. We've taught our code to make logical decisions. But we still haven't built the "safety net" for when it receives bad data (like the ValueError). That is the final piece of our foundation. Tomorrow, we build the safety net: Error Handling. #Python #DataScience #SoftwareEngineering #AI #Developer #Logic

🚀 ThinAgents, built on top of LiteLLM (YC W23), now introduces WebUI! You can now instantly spin up a web interface to chat with your AI agents — directly from your Python script. Define your agents and tools as usual, run the file, and boom, a clean, interactive UI opens up to talk to your agent in real time. No setup. No frontend code. No config. Just pure plug-and-play GenAI magic WebUI is still very early, and I’ll soon be adding: 🧠 Session management 🪣 Memory 🔗 MCP integrations all within the UI install now using `pip install -U thinagents` GitHub: https://lnkd.in/g5RpFZqT Excited to see where this goes and what people build with it! #AI #GenAI #Python #ThinAgents #LiteLLM #LangGraph #SvelteKit #FastAPI

AI ≠ Productivity 🤖 AI just cost me an hour of debugging time. I was fixing a bug and turned to GitHub Copilot for help. It removed one line and added 56 lines of new logic. It looked complicated but “smarter,” so I trusted it. Ten minutes later, the codebase exploded with new errors — each one a consequence of that over-engineered “fix.” When I stepped back and debugged it myself, the real solution was just one line and two minor edits. 💡 Lesson learned: AI can write code fast, but debugging still demands understanding. This is part of my 'build-in-public' series for stocksimpy, a Python stock backtesting library that I built from scratch. Check out the repo and make sure to follow me for more updates: https://lnkd.in/eu5UyyWA

𝑰𝒎𝒂𝒈𝒊𝒏𝒆 𝒉𝒂𝒗𝒊𝒏𝒈 𝒂𝒏 𝒂𝒔𝒔𝒊𝒔𝒕𝒂𝒏𝒕 𝒘𝒊𝒕𝒉 𝑳𝑳𝑴 𝒑𝒐𝒘𝒆𝒓. You give it a role, choose LLM model according to your preference, ask a query, and it brings back exactly what you need — structured, simplified, and crystal clear answer. For the past couple of weeks, I’ve been exploring how to actually build something like that — from scratch. 𝑯𝒆𝒓𝒆’𝒔 𝒘𝒉𝒂𝒕 𝑰’𝒗𝒆 𝒃𝒆𝒆𝒏 𝒘𝒐𝒓𝒌𝒊𝒏𝒈 𝒘𝒊𝒕𝒉 👇 ➡️ Python – the clean base to bring everything together ➡️ LangChain + LangGraph AI – to shape how the system thinks, reasons, and connects data ➡️ Pydantic – keeps the data organised and structured (absolute lifesaver for validation) ➡️ Tavily – adds real-time, relevant context from the web ➡️ OpenAI + Gemini – for responses that actually make sense ➡️ FastAPI + Streamlit + React Agent – to bring it all to life with a proper interface This whole process has been less about “coding” and more about understanding how systems can think, reason, and act. And honestly — it’s been fascinating. I’ll share a breakdown of why each library is used and how they all work together — from back-end flow to front-end interaction. 𝑪𝒉𝒆𝒄𝒌𝒐𝒖𝒕 𝑾𝒐𝒓𝒌𝒊𝒏𝒈 𝑽𝒊𝒅𝒆𝒐 👇 #Python #LangChain #LangGraph #Pydantic #Tavily #FastAPI #Streamlit #React #Gemini #OpenAI #FullStackDev #LearningInPublic #ProductManagement #TechInnovation #DevelopersJourney #BuilderMindset #Engineering

This weekend, I built a small technical project for fun: a lightweight local vector engine (shards, HNSW indexing, fallback search, modular embedder, API server, logs…). Nothing commercial. Just something I decided to open-source for the community. Not perfect, not “enterprise-grade”. Just clean, simple, functional code — the kind of tool I couldn’t find anywhere in a minimal form. If it helps someone in their experiments (RAG, LLMs, embeddings, research…), good. If not, that’s fine too. Here is the open-source release: 🔗 https://lnkd.in/guK-Qknz Use it, break it, learn from it, improve it. Sharing what I build is my way to give back. #Python #AI #LLM #OpenSource #MachineLearning #Developer #RAG

Keeping with the promise of providing open access, I’m sharing an AI Agent skeleton that I developed using LangChain and Flask. It is a lightweight framework for building more complicated AI agents, LLM workflows, and intelligent web APIs. (An example is included in the code which you can run right away on a local server!) This open-access repo has modular chains, tools, and memory components, making it easy for anyone to learn and even customize, extend, and deploy your own agent within minutes. It also comes with an instruction set to understand how each part works. 🌐 Check it out: https://lnkd.in/dcAn9kw2 The question; How would you use it? and if you have ideas, let's chat! #AI #LangChain #Flask #OpenSource #LLM #Python #AIagents

Excited to share my project, the "AI-Powered Timed MCQ System"! 🚀 Its main feature? 💡 You just enter any topic imaginable, and the system instantly generates a challenging, 15-question timed quiz. It's a full-stack web app 💻 designed to replace static, boring quizzes with a dynamic, real-time assessment tool. Here's how it works: Client (React/TS) 👨💻: User inputs a topic (e.g., "Solar System," "Python Decorators"). Server (FastAPI) 🐍: My Python backend securely calls the Google Gemini API. AI (Gemini) ✨: Generates a 15-question JSON array categorized by difficulty (Simple, Moderate, Difficult). Client (React) ⏱️: Renders the quiz interface, managing a 5-minute overall timer and per-question timers (10, 20, or 30 seconds). Results 📊: A final, difficulty-weighted score is presented to the user. This project was a great deep-dive for me into full-stack development and practical Generative AI integration. 🧠 #FullStack #AI #GenerativeAI #Python #FastAPI #React #EdTech #ProjectDemo #SoftwareDeveloper

🧠 Built a Fully Local RAG, No API Keys Needed I’ve added a new feature to my Lumini AI project, a fully local RAG system that works without relying on any external APIs. I wanted to deeply understand how RAG actually works, not just call a prebuilt API. So I built it from the ground up and integrated it into Lumini, which lets users “Ask your repo” questions. 🔹 Fetches repo files directly from GitHub 🔹 Generates embeddings locally using Sentence Transformers 🔹 Stores them in Qdrant for vector search 🔹 Retrieves the most relevant context 🔹 Generates answers with Llama 3.1 7B running locally 🔹 Streams clean, markdown-formatted responses to the frontend No API keys. No recurring costs. Just open-source intelligence - fast, private, and fully integrated into Lumini. What do you think of this? What other features or experiments would you try with a local RAG setup like this? 🤔 #AI #RAG #LLM #MachineLearning #Llama3 #FastAPI #Python #Developer #BuildInPublic #AIEngineering #MLOps #FullStackAI

Most engineers have used an MCP server. Few have built one. In this new post, Pablo Tomas Fernandez walks through creating a fully working MCP server that connects directly to Semaphore’s API. It takes under 50 lines of Python to make your CI/CD setup conversational — list projects, pull build data, even ask why a build failed. A short, hands-on look at what AI + CI/CD actually feels like. Read the full post → https://lnkd.in/dCtjmgDX