Muhammad Moiz — Writing https://www.muhammadmoiz.com Essays, TILs, build logs, and notes from an AI/ML engineer in Paris. en I Panicked in a Container With No Curl Until I Remembered Bash Has a Secret HTTP Client https://www.muhammadmoiz.com/blog/34-bash-http https://www.muhammadmoiz.com/blog/34-bash-http No curl, no wget, no nc in a 12 MB container: bash's /dev/tcp redirection let me hand-write an HTTP request and hit the health endpoint. Tue, 16 Jun 2026 00:00:00 GMT tutorial My AI Refused to Use Markdown. Turns Out It Was Secretly Following Orders I Never Gave It. https://www.muhammadmoiz.com/blog/31-hermes-agent-formats https://www.muhammadmoiz.com/blog/31-hermes-agent-formats My Markdown renderer was not broken. Hermes Agent was quietly telling the model to answer in plain text because I connected over HTTP. Tue, 16 Jun 2026 00:00:00 GMT notes I Designed a Multi-Model System. My Server Room Smelled Like Burnt Coffee. https://www.muhammadmoiz.com/blog/37-multi-model-design https://www.muhammadmoiz.com/blog/37-multi-model-design Five patterns cover almost every multi-model setup: sequential, router, parallel, hierarchical, ensemble, and each one bills you differently. Mon, 15 Jun 2026 00:00:00 GMT deep-dive The Little Share Buttons That Couldn't https://www.muhammadmoiz.com/blog/27-share-buttons https://www.muhammadmoiz.com/blog/27-share-buttons Three studies, three audiences, the same number: about 0.2% of visitors ever touch a share button. Everyone else just copies the link. Mon, 15 Jun 2026 00:00:00 GMT essay I Used One Model for Everything. My Wallet Cried. https://www.muhammadmoiz.com/blog/18-model-routing https://www.muhammadmoiz.com/blog/18-model-routing Running one big model for every task burns money and latency, so I started routing by capability, cost, and speed instead. Mon, 15 Jun 2026 00:00:00 GMT essay I Broke My LLM in Production So You Don't Have To: A Guardrails Story https://www.muhammadmoiz.com/blog/17-llm-guardrails https://www.muhammadmoiz.com/blog/17-llm-guardrails An hour after deploy, a user asked my chatbot something it should never have answered, and that is how I learned guardrails are the product. Mon, 15 Jun 2026 00:00:00 GMT tutorial I Miss When Claude Was Boring https://www.muhammadmoiz.com/blog/24-claude-personality https://www.muhammadmoiz.com/blog/24-claude-personality Somewhere between safety guardrails and anti-sycophancy training, my chatbot turned into a debate opponent. Sun, 14 Jun 2026 00:00:00 GMT essay How "Load-Bearing" Became the New "Delve" https://www.muhammadmoiz.com/blog/14-load-bearing-speak https://www.muhammadmoiz.com/blog/14-load-bearing-speak Claude's favourite adjective is leaking out of code reviews and into everyone's vocabulary, and it is going the same way as 'delve'. Sun, 14 Jun 2026 00:00:00 GMT essay The AI Assistant Architecture That Took Me Six Months and a Thousand Dollars in API Bills to Figure Out https://www.muhammadmoiz.com/blog/10-ai-assistant-architecture https://www.muhammadmoiz.com/blog/10-ai-assistant-architecture Five layers (LLM, memory, tools, routing, observability), the failure modes that ambushed me, and the boring stack I actually run today. Thu, 11 Jun 2026 00:00:00 GMT deep-dive I Am the Bottleneck https://www.muhammadmoiz.com/blog/13-human-bottlenecks https://www.muhammadmoiz.com/blog/13-human-bottlenecks AI keeps promising a 10x me, but executive function, intelligence, and knowledge all live inside my head where no amount of scaffolding reaches. Mon, 18 May 2026 00:00:00 GMT essay The One llama.cpp Feature That Almost Made Me Write a Polite Angry Email https://www.muhammadmoiz.com/blog/28-llamacpp-unload https://www.muhammadmoiz.com/blog/28-llamacpp-unload llama.cpp's router has no unload-all endpoint, so I built one out of curl, jq, and a while loop. Six lines, and it beats a button. Sat, 16 May 2026 00:00:00 GMT tutorial PKM, RAG, Wikis, and Memory Systems Are Not the Same Thing (And Yeah, I Got This Wrong Too) https://www.muhammadmoiz.com/blog/11-pkm-rag-comparison https://www.muhammadmoiz.com/blog/11-pkm-rag-comparison PKM, wikis, RAG, and AI memory all store information, but each one optimizes a different layer of the stack, and forcing one to do every job ends badly. Wed, 13 May 2026 00:00:00 GMT essay I Have 14 Note-Taking Apps and Remember Nothing: A Confession https://www.muhammadmoiz.com/blog/32-km-overview https://www.muhammadmoiz.com/blog/32-km-overview Fourteen note apps later, what survived was not a tool but a few methods: atomic notes, an actionability filter, and links I will actually follow. Sun, 19 Apr 2026 00:00:00 GMT essay How I Quit Ollama Cold Turkey (And Why My Laptop Runs Faster Now) https://www.muhammadmoiz.com/blog/01-ollama-alternative https://www.muhammadmoiz.com/blog/01-ollama-alternative Six months of Ollama, then I read the GitHub issues: what the wrapper was hiding, and the llama.cpp setup I run instead. Wed, 15 Apr 2026 00:00:00 GMT essay I Almost Bought a $300 Gadget. Then My Apple Watch Called Me an Idiot. https://www.muhammadmoiz.com/blog/23-apple-watch-transcription https://www.muhammadmoiz.com/blog/23-apple-watch-transcription The recording hardware was already on my wrist: Voice Memos, iCloud sync, and launchd replaced the watchOS app I never had to build. Sun, 29 Mar 2026 00:00:00 GMT essay Can AI coding agents relicense your GPL code without your permission? https://www.muhammadmoiz.com/blog/15-coding-agents-relicense https://www.muhammadmoiz.com/blog/15-coding-agents-relicense A real fight over chardet 7.0.0 has the open source world asking whether LLMs enable a modern kind of clean room reimplementation. Thu, 05 Mar 2026 00:00:00 GMT essay What I Learned From Taking a Self-Hosted AI Assistant Seriously https://www.muhammadmoiz.com/blog/19-self-hosted-ai-assistant https://www.muhammadmoiz.com/blog/19-self-hosted-ai-assistant Running a model is infrastructure work, designing an assistant around it is systems work, and confusing the two is where most local AI projects stall. Tue, 03 Mar 2026 00:00:00 GMT essay I let a local LLM read my private journal. I was hoping for surprises. https://www.muhammadmoiz.com/blog/22-llm-journal-analysis https://www.muhammadmoiz.com/blog/22-llm-journal-analysis I fed 193,761 words of journal entries to a local LLM and got back a prettier version of my own memory. Fri, 13 Feb 2026 00:00:00 GMT essay The Local LLM Hosting Odyssey: What I Learned Running Models Every Way Imaginable https://www.muhammadmoiz.com/blog/16-llm-hosting https://www.muhammadmoiz.com/blog/16-llm-hosting I ran local models every way I could find, from Ollama to llama.cpp to vLLM to cloud APIs, and here is what each one is actually good for. Thu, 12 Feb 2026 00:00:00 GMT deep-dive Write Badly and the World Moves On https://www.muhammadmoiz.com/blog/33-write-badly https://www.muhammadmoiz.com/blog/33-write-badly A 45 minute timer, cheap rewrites, and publishing the flawed thing beat years of sanding every sentence down to a nub. Fri, 06 Feb 2026 00:00:00 GMT essay I Installed Anaconda, Miniconda, Mamba, and Conda (They Are Not the Same Thing) https://www.muhammadmoiz.com/blog/25-conda-comparison https://www.muhammadmoiz.com/blog/25-conda-comparison Anaconda, Miniconda, conda, and Mamba are not competing products, they are layers in a stack, and picking the wrong combo costs you gigabytes. Wed, 28 Jan 2026 00:00:00 GMT deep-dive The LLM search engine that lives in my pocket (and takes 17 minutes to answer anything) https://www.muhammadmoiz.com/blog/07-llm-web-search https://www.muhammadmoiz.com/blog/07-llm-web-search I pay twenty dollars a month for a machine that reads hundreds of web pages, writes a report I skim, and hands me the links I actually wanted. Wed, 03 Dec 2025 00:00:00 GMT essay The Neural Esperanto: What Happens When You Make AI Models Speak the Same Language https://www.muhammadmoiz.com/blog/21-cross-modal-embeddings https://www.muhammadmoiz.com/blog/21-cross-modal-embeddings Contrastive learning shoves images and text into one shared vector space, and the engineering around that space is where the real work lives. Fri, 21 Nov 2025 00:00:00 GMT deep-dive I Built a Web Search MCP Server in Python (And Only Cried Twice) https://www.muhammadmoiz.com/blog/20-mcp-server-python https://www.muhammadmoiz.com/blog/20-mcp-server-python Building an MCP server in Python with DuckDuckGo search and Playwright scraping, plus the rate limits and caching bugs I hit on the way. Sun, 26 Oct 2025 00:00:00 GMT tutorial My Conda Era: A Python Environment Confession https://www.muhammadmoiz.com/blog/29-pyenv-alternative https://www.muhammadmoiz.com/blog/29-pyenv-alternative I hoarded 47 Conda environments until the licensing bill and the disk usage caught up with me. pyenv was the intervention I needed. Thu, 23 Oct 2025 00:00:00 GMT essay I was a Python environment victim (and then uv showed up) https://www.muhammadmoiz.com/blog/04-uv-python-tool https://www.muhammadmoiz.com/blog/04-uv-python-tool A decade of pyenv, venv, pip and poetry duct-taped together, collapsed into two commands, plus the uvx trick I use daily. Thu, 23 Oct 2025 00:00:00 GMT essay I Have Tried Every Note-Taking App So You Do Not Have To (A PKM Tool Guide) https://www.muhammadmoiz.com/blog/09-pkm-part2 https://www.muhammadmoiz.com/blog/09-pkm-part2 Obsidian, Notion, Roam, Logseq, Evernote, TiddlyWiki: what each one is actually good for, and the setup I stopped tinkering with. Mon, 14 Jul 2025 00:00:00 GMT tutorial Your Brain Is Leaking: Why You Need a Second Brain (and How to Stop Forgetting Everything) https://www.muhammadmoiz.com/blog/09-pkm-part1 https://www.muhammadmoiz.com/blog/09-pkm-part1 I tried every note-taking app there is and forgot most of what I wrote, so here is what personal knowledge management actually needs to do. Mon, 14 Jul 2025 00:00:00 GMT essay I Built an MCP Server in Go So You Don't Have To https://www.muhammadmoiz.com/blog/12-mcp-server-go https://www.muhammadmoiz.com/blog/12-mcp-server-go Building a Model Context Protocol server in Go: the SDK situation, the initialization handshake, and the JSON-RPC gotchas that cost me hours. Fri, 11 Jul 2025 00:00:00 GMT tutorial The Python Package Manager That Finally Made Me Stop Yelling at My Terminal https://www.muhammadmoiz.com/blog/26-uv-guide https://www.muhammadmoiz.com/blog/26-uv-guide uv replaces pip, virtualenv, pyenv, pip-tools, and pipx with one Rust binary, and I have not touched pip directly since I switched. Thu, 26 Jun 2025 00:00:00 GMT tutorial Fifty Ways to Trick Yourself Into Being a Functional Adult https://www.muhammadmoiz.com/blog/02-adhd-systems-part2 https://www.muhammadmoiz.com/blog/02-adhd-systems-part2 The daily tactics: Todoist as a brain prosthetic, inbox bankruptcy, visual field management, and the productivity traps I keep falling into. Thu, 12 Jun 2025 00:00:00 GMT essay The Voltage Theory: What I Learned After I Stopped Trying to Out-Will My Own Brain https://www.muhammadmoiz.com/blog/02-adhd-systems-part1 https://www.muhammadmoiz.com/blog/02-adhd-systems-part1 Why medication is the prerequisite tech, why energy behaves like voltage rather than spoons, and three procrastinations with three different fixes. Thu, 12 Jun 2025 00:00:00 GMT essay My Favorite Object Detection Framework Died and I Didn't Even Get to Say Goodbye https://www.muhammadmoiz.com/blog/35-mmdetection https://www.muhammadmoiz.com/blog/35-mmdetection MMDetection stopped moving in early 2024 and nobody announced it, which is what happens when a framework rests on one institution. Fri, 04 Apr 2025 00:00:00 GMT essay The Hockey Stick in My Chat History https://www.muhammadmoiz.com/blog/03-claude-workflow https://www.muhammadmoiz.com/blog/03-claude-workflow I exported a year of chat history and graphed it: words per day went up 727%, and here is what I was actually talking about. Wed, 01 Jan 2025 00:00:00 GMT essay When the Reward Signal Is Also an AI That Can Be Gaslit: Reward Hacking Part 2 https://www.muhammadmoiz.com/blog/08-reward-hacking-part2 https://www.muhammadmoiz.com/blog/08-reward-hacking-part2 What happens when the reward signal is itself a model: overoptimization, sycophancy, biased LLM graders, and agents that rewrite their own reward. Thu, 28 Nov 2024 00:00:00 GMT deep-dive When You Tell a Robot to Get Points and It Becomes a Gremlin: Reward Hacking Part 1 https://www.muhammadmoiz.com/blog/08-reward-hacking-part1 https://www.muhammadmoiz.com/blog/08-reward-hacking-part1 Reward hacking, from the boat that races in circles to the robot hand that only pretends to grab things, and why smarter agents make it worse. Thu, 28 Nov 2024 00:00:00 GMT deep-dive I Tried to Build Midjourney Blend at Home and Got a Diffusion Smoothie Instead https://www.muhammadmoiz.com/blog/36-image-mixer-diffusion https://www.muhammadmoiz.com/blog/36-image-mixer-diffusion I fed random CLIP crops of two images into a fine-tuned diffusion model hoping for compositional blending and got a smoothie. Fri, 19 Apr 2024 00:00:00 GMT notes I Built a Diffusion Model That Could Copy Anyone's Face. Then I Lost It Like a Pair of Sunglasses. https://www.muhammadmoiz.com/blog/30-face-mixer-diffusion https://www.muhammadmoiz.com/blog/30-face-mixer-diffusion I trained a diffusion model that clones a face from one photo, got face and background mixing working, then lost the checkpoint forever. Fri, 19 Jan 2024 00:00:00 GMT deep-dive My Agent Phase, and Why I Got Out https://www.muhammadmoiz.com/blog/06-llm-agents-thoughts https://www.muhammadmoiz.com/blog/06-llm-agents-thoughts I built a few LLM agent systems, watched more of them fail, and came out convinced the general cognitive architecture is the wrong target. Fri, 05 Jan 2024 00:00:00 GMT essay What Happens When You Actually Run LLM Agents in the Wild https://www.muhammadmoiz.com/blog/05-llm-agents-part2 https://www.muhammadmoiz.com/blog/05-llm-agents-part2 What HuggingGPT, ChemCrow, generative agents and AutoGPT actually do in practice, and the four walls every agent still runs into. Fri, 23 Jun 2023 00:00:00 GMT deep-dive I Tried Building an LLM Agent and Accidentally Gave It a To-Do List It Could Never Finish https://www.muhammadmoiz.com/blog/05-llm-agents-part1 https://www.muhammadmoiz.com/blog/05-llm-agents-part1 The three things you bolt onto an LLM to make it an agent: planning, memory and tool use, and where each one starts leaking. Fri, 23 Jun 2023 00:00:00 GMT deep-dive