A step-by-step guide that shows how to create your own lightweight feature store MCP server from scratch using FastMCP, run it through uv, and integrate it with Claude Desktop. This is a practical example of building a useful MCP server that ML engineers can actually use.
Open-source LLMs are shipped in two axes of variation:
-Instruct, -Distill, -A3B, …) tell you how the checkpoint was produced and what it's good at..gguf, .gptq, …) tells you how the weights are packed for inference on different hardware.Understanding both axes lets you avoid downloading 20 GB for nothing or fighting CUDA errors at 3 a.m.
Running generative-AI models on-device means zero cloud costs, no upload limits, and full control of your checkpoints. Below is a quick guide to five of the most popular macOS-ready front-ends and launchers.
This guide compares the five most popular local toolchains, complete with download links, quick overviews, and pros & cons. A comparison table follows for easy reference.
pyproject.tomlThink of pyproject.toml as the package.json for Python. Whether you prefer .venv, pyenv, or uv, putting all your project's metadata, dependencies, and tooling into one tidy TOML file simplifies development and boosts collaboration.
~/.zprofile → one-shot, login-shell initialization (think "environment/bootstrap") 🔧~/.zshrc → every interactive prompt (think "daily driving experience") 🎮Use both in tandem: keep your environment reliable with ~/.zprofile, and your shell pleasant and tweakable with ~/.zshrc ✨
The field of machine learning has undergone a seismic shift with the rise of large-scale foundation models - from giant language models (LLMs) like GPT-4 to image diffusion models like Stable Diffusion. As a result, the way we build and operate ML systems (MLOps) looks very different today than it did just a few years ago. In this post, we'll explore how ML infrastructure and MLOps practices have evolved - contrasting the "classic" era of MLOps with the modern paradigms emerging to support foundation models. We'll highlight what's changed, what new patterns and workflows have emerged.
The race to build bigger, better language models continues at breakneck speed. Today's state-of-the-art models require massive computing resources that no single GPU can handle. Whether you're training a custom LLM or deploying one for inference, understanding how to distribute this workload is essential.
This guide walks through practical strategies for scaling LLMs across multiple GPUs and nodes, incorporating insights from Hugging Face's Ultra-Scale Playbook.
Here's my distilled, 10‑step workflow to transform a vanilla macOS install into a ready to-go AI engineering working station.