How an engineer searches a codebase: A deep dive

🦀 What’s stopping you from learning Rust? You’ve got Codex, Claude Code, and half a dozen AI copilots that can literally explain every line as you type. You can build web backends, APIs, data pipelines, and even neural interfaces — all in one language that doesn’t panic when things get fast. So be honest: 👉 Do you not like going fast? 👉 Do you not like being secure? 👉 Or are you still waiting for permission to try something that doesn’t segfault your weekend? Rust isn’t just a “systems language” anymore. It’s a full-stack AI language if you know where to look. Here’s what that stack looks like: 🧠 AI + ML: tch-rs (PyTorch bindings), ort (ONNX Runtime), burn, linfa, fastembed. ⚙️ APIs & Backends: axum, actix, poem, warp. 💾 Data Layer: polars, arrow, duckdb, sqlx. 🎨 Front-end / WASM: leptos, yew, dioxus. 🤖 AI Integrations: use Codex, Claude Code, or local llama.cpp bindings to wire inference directly into your stack. Imagine building an app that fetches data, trains a model, serves inference through Axum, and compiles to WASM for a browser dashboard — all in Rust. That’s not the future. That’s just today, if you stop hesitating and let AI guide you while you learn. Rust gives you: Memory safety without a GC. Zero-cost abstractions. Predictable performance. Type safety that feels like armor plating. And now, with AI copilots writing boilerplate and explaining lifetimes, the biggest barrier to entry is gone. So… what’s stopping you? 🦀 #Rust #AI #FullStackDevelopment #MachineLearning #Codex #ClaudeCode #WebAssembly #Axum #Burn #Polars #ONNX #DevCommunity

This is some of the best advice for operationalizing successful AI. 3 minute read and everything in the article is golden. Real world: for the last year, we tried to make SML's come true on AIPCs and we ran across at least 6 of these points. We muscled through them, but it was exactly that - one hard knock at a time, one after the other. My favorite quote from the article: "Amateurs talk about models. Professionals talk about data pipelines.”. Because in the end, it's all about the data. And maintainability. https://lnkd.in/eykGij5s

PyTorch is Not Just Research Anymore: It’s a Production Powerhouse! When PyTorch first gained traction, it was celebrated for its Pythonic flexibility and dynamic computational graph, making it the undisputed champion of AI research. But today, the story is much bigger. With the introduction of PyTorch 2.0 and the torch.compile() API, the framework has fundamentally changed how it executes code under the hood, making it a serious production contender. The secret sauce is the new compiler stack (TorchDynamo, AOTAutograd, TorchInductor), which gives us the best of both worlds: Eager Mode Flexibility: You still get the easy debugging and rapid iteration of the Python-first, define-by-run style. Graph Mode Performance: torch.compile() safely converts your model code into optimized computational graphs, generating highly efficient kernels (often using Triton) to deliver massive speedups in training and inference. This means we can prototype cutting-edge models quickly and deploy them at scale efficiently, without painful code conversion. If you’re not using torch.compile() yet, you are leaving serious performance on the table. It's time to leverage the compiler magic and experience the true power of PyTorch! #PyTorch #DeepLearning #MachineLearning #AI #TorchCompile #TechInnovation #MLOps

What I learned from processing 5M+ documents with RAG: it's not just about the tech stack. • Query Generation: LLM-generated semantic + keyword queries improved results by covering a larger surface area. • Reranking: 5 lines of code that shifted chunk ranking more than expected. Takeaway: RAG success depends on a combination of query generation, reranking, and chunking strategy. #rag #llm #nlp

Just finished Google's "Advanced: Generative AI for Developers" course. Time to put theory into practice 👇 I built DSM (Damage Segmentation System) - an AI that detects and localizes car damage in real-time. Here's what I did: 🔧 Started with open-source YOLO11m weights (CarDD dataset) → Fine-tuned on additional scratch/dent samples using transfer learning → Applied data augmentation (rotation, lighting, perspective shifts) to improve edge cases → Result: 91% confidence on real-world damage detection 💡 But detection alone isn't enough for real use. So I added a second AI layer: → Integrated Gemini 2.5 Pro for spatial reasoning → Crops detected region + context → Returns precise automotive part identification ("Front Left Fender" not "left side") → Uses industry-standard perspective (driver's POV, not camera angle) 🎯 Final System: • 6 damage types: scratches, dents, cracks, glass, lights, tires • Real-time inference on CPU • Before/after comparison mode • Production-ready Gradio interface 🛠️ Tech Stack: PyTorch • YOLO11m • Gemini 2.5 Pro • OpenCV • Gradio The two-stage approach (detection + localization) is what makes this actually useful for insurance claims and service centers. Theory → Practice → Production. Want to learn more about DSM or explore other AI innovations I'm working on? Feel free to reach out, or visit my website: https://virelio.nl/ What would you build with generative AI? #AI #ComputerVision #MachineLearning #GenerativeAI #PyTorch #Google #InsurTech #Python #BuildInPublic

Everyone talks about AI, but only a few build it into something that actually works for people. Over the last year, I’ve been working on production-grade RAG (Retrieval-Augmented Generation) systems — assistants that connect large language models with private knowledge bases. Instead of “chatbots that hallucinate,” these systems give grounded, explainable answers based on verified company data. One of my recent projects involved building a nutrition assistant (“Talia”) that analyzes user messages, retrieves relevant knowledge from an Outline-based knowledge base via LlamaIndex, and responds in a warm, human tone — all in real time. The backend stack: FastAPI + Supabase + OpenAI Assistants API, orchestrated through n8n for external integrations. What I’ve learned: - The magic isn’t just in GPT — it’s in how you structure and connect the data. - A good RAG pipeline is 80% engineering, 20% prompt design. - “Personalized AI” starts when context retrieval becomes reliable and fast. - Hybrid stacks (Python + no-code) can ship AI faster than many expect. AI in 2025 isn’t about experiments anymore — it’s about infrastructure.

Been deep into my new project lately, using vibe coding with an AI-powered IDE. It honestly feels like pairing up with a co-pilot that sometimes needs you to fly the plane. But let me remind you, we still need to know the syntax before giving semantic orders. If you don’t know which function maps to which operation, no LLM can read your mind. A one-liner like “I want this, I want that or optimize my code” means nothing unless you define your logic and architecture. These AI IDEs are brilliant at pattern recognition, but they also hallucinate. You need to describe your style, even your frequency if you want the output to match your intent. You’ve got to catch the drift, debug not just the code, but the reasoning behind the model’s choices. I’ve always thought of it less like refactoring code and more like fine-tuning a neural net mid-run: that is precision prompts, quick iteration loops, and error-handling instincts matter more than ever. It’s debugging 2.0, you’re no longer just debugging lines of code, because you’re debugging the vibe itself. The compiler is Intelligent, but you’re still the brain steering the flow. #ArtificialIntelligence #AI #VibeCoding #GenerativeAI

My first autonomous AI agent - and it actually works 🚀 I spent the last week learning about agentic AI by building something hands-on. The concept: Instead of hardcoding "if this, then that" logic, give an AI access to tools and let it decide what to do. My test case: "Draw a rectangle and show the sum of 50+75 inside" What the AI figured out on its own: 1. "I need a canvas first" → calls open_drawing_app() 2. "Now draw the rectangle" → calls draw_rectangle() 3. "Calculate the sum" → calls add(50, 75) 4. "Display the result" → calls add_text_in_drawing_app("125") I didn't program this sequence. The AI reasoned through it. Technologies I learned: - Model Context Protocol (MCP) - lets AI discover and use tools - Google Gemini for the reasoning engine - Python for building the tool server - 22 custom functions the AI can call Biggest challenges: - Getting Mac permissions right - Understanding MCP's stdio protocol - Debugging when screen coordinates didn't match canvas size - Realizing the AI needs feedback from each step This opened my eyes to what's possible beyond chatbots. Same pattern could automate real workflows - data analysis, report generation, system monitoring. Code is on GitHub if you want to learn alongside me: https://lnkd.in/dUVy9FeN Check the demo here : https://lnkd.in/d234rYkd For those building agents - what was your "aha moment"? #AgenticAI #MCP #Python #AILearning.

Imagine an AI that doesn’t just respond but acts—handling tasks like answering customer queries, managing finances, or even tutoring students. That’s Agentic AI, and it’s changing the game! In my latest video, I break it all down: ✅ What is Agentic AI, and how does it work? ✅ Real-world use cases transforming industries. ✅ Step-by-step guide to building your own Agentic AI in vanilla Python—no fancy libraries, just pure logic and innovation! 💡 Ready to dive into the next big thing in AI? 🎥 Watch Now Let’s explore the future of autonomous systems together! Video Link: youtu.be/dsbjMbK_9Gw   Channel Link: youtube.com/@datatrek