Ismail Canga’s Post

I published a new article about something many developers will appreciate: how to recover your Claude Code chat history in VS Code without losing the early context that gets compacted away. The key idea is simple: - Claude Code stores history locally in ~/.claude/projects. - The logs are saved as JSONL files. - Context compaction affects the live session, not the stored history. - A Python CLI can help you retrieve and read the transcripts. If you’ve ever lost a good idea because you clicked “yes” too quickly or forgot to save your prompt trail, this workflow can save you time and frustration. Read the article here: https://lnkd.in/epUjEF6f #ClaudeCode #VSCode #Python #DataScience #AI #DeveloperProductivity #Automation #Substack

You don't need LangChain to build AI agents. You need to understand what's inside it. I spent the last month building agents from scratch just Python, Ollama, and the ReAct loop. No frameworks. No abstractions. Nothing I couldn't debug. 6 concepts kept coming back: → Anatomy — the 4 parts of every agent → The loop — how decisions actually happen → Tools — what they really are (not magic) → Memory — the illusion the harness creates → Scaling — what breaks at each stage → Orchestration — when one agent isn't enough I put them in a visual guide. 11 slides. Save it. Open it next time your framework does something weird. Share it with someone building their first agent. This is the reference I wish I had on day 1. 👇 Swipe through. #AIAgents #AIEngineering #LLM #MachineLearning #SoftwareEngineering #Python #BuildInPublic #FirstPrinciples #TechCareers

Wild to see Claude Code, one of the most talked-about AI coding tools right now, go viral like this. v2.1.88 was accidentally shipped with a debug/source map file, exposing a large part of the internal TypeScript codebase. Anthropic said it was a release mistake caused by human error, not a hack, and no customer data was leaked. The craziest part? The internet moved so fast that people were already digging through it and rebuilding ideas around it in Python and Rust. Just shows how one bad release can put your product everywhere in a few hours. checkout python implementation https://lnkd.in/gWEV_XKc from instructkr. #ClaudeCode #Anthropic #AI #SoftwareEngineering

Built a voice controlled AI agent using Python, Gradio, and Groq APIs (Whisper + Llama 3.3). Implemented a 3-tier intent classifier (regex, keyword scoring, LLM fallback) that minimizes API calls. Supports 4 intents: create file, write code, summarize, and general chat. All file operations sandboxed to output/ directory. layering features: compound commands, human-in-the-loop confirmation, session memory, and graceful error handling.

LLM apps are powerful, but they're also hard to debug and expensive to run. Observability shouldn't make that worse. In our latest blog, we break down how the Python OpenTelemetry SDK impacts performance in LLM workloads and what happens when you rethink the stack. If you're building with LLMs and care about performance, this is worth a read 👉 https://lnkd.in/e_aHurKw #honeycomb #OpenTelemetry #AI #LLMs

🚀 Built my first RAG (Retrieval-Augmented Generation) Chatbot using Python! Instead of guessing, this chatbot reads, understands, and answers directly from custom data 📄➡️🤖 Powered by FAISS, HuggingFace embeddings, and Groq LLM, it delivers fast and context-aware responses. 💡 From static text → to intelligent conversations This is a small step into the world of AI-powered applications, but a big leap in how machines interact with knowledge. #AI #MachineLearning #LangChain #Python #RAG #GenAI #DataScience

A few weeks ago I couldn't write a single line of code. Tonight I built a Python script that: → Reads my own Discord log files → Sends them to Claude AI via API → Gets back a real-time status report about my progress The script is called Report.py. It runs locally, pulls data from my Realtime Discord Logger, and uses an LLM to summarize what's happening in plain English. No tutorial. No copy-paste. Just debugging until it worked. This is what Phase 2 looks like — connecting real data to real AI. Full project on GitHub 👇 https://lnkd.in/eCwq_3XZ #AIAutomation #Python #BuildingInPublic #CareerChange

Built a GenAI application that converts YouTube videos into structured articles, downloadable PDFs, and responsive webpages. The system uses a multi-step pipeline: transcript extraction, content cleaning, article generation, and multi-format output creation. Tech stack: Python, Streamlit, LangChain, Groq (LLaMA 3.3), FPDF This project helped me understand how to design end-to-end GenAI workflows beyond simple summarization. 🔗 GitHub: https://lnkd.in/gygZUgGG #Innomaticsresearchlabs #GenerativeAI #Langchain #Python #Datascience

Standard LLM chains are great for straight lines, but real-world problems are messy. They need loops, second guesses, and human check-ins. LangGraph solves that problem. It’s the "pro mode" for LangChain. I just put together a quick post on how to get started with this framework with a simple code example. Here’s the TL;DR: ✅ Define State Early: Use TypedDict to give your agent a reliable memory. ✅ Conditional Edges: Let the LLM decide when to pivot or retry. ✅ Persistence: Never lose progress again with built-in checkpointers. ✅ Human-in-the-Loop: Add "pause" buttons for high-stakes actions. ✅ Keep Nodes Pure: Small, focused functions make debugging a breeze. Check out the full breakdown below! 👇 https://lnkd.in/e_T-bCaG #AI #DataScience #Python #LangChain #LangGraph #MachineLearning #LLM

𝐋𝐋𝐌𝐬 𝐝𝐨𝐧'𝐭 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐝𝐨 𝐚𝐧𝐲𝐭𝐡𝐢𝐧𝐠.   They generate text. So when an “AI agent” queries a database, sends an email, or runs a command — something else is doing the real work. An orchestration layer most people never see. I wrote a breakdown of what that layer actually is — the 𝐑𝐞𝐀𝐂𝐓 𝐥𝐨𝐨𝐩 behind agent frameworks — built from scratch in ~40 lines of Python, with an interactive stepper to watch a full run end-to-end. Check out my blog post about this subject: → https://lnkd.in/e3K94--D #AIAgents #AgenticAI #LLM #AIEngineering #SoftwareEngineering