Why this exists

There has never been more Python AI tooling, and picking the right library for a real project has never been harder. Those two things are related.

PyStack is a curated map of the tools I'd actually use if you handed me a project tomorrow. Organized by where each tool sits in the stack, from infrastructure to observability. Opinionated by design.

The problem

Open PyPI and search for anything AI-related. You'll find hundreds of packages, most of them recent. By early 2026, there were more than 700 AI-related packages on PyPI. Around 40 percent hadn't shipped a release in the past twelve months. A lot of what's left is thin wrappers around someone else's API, or demos that never made it past the README.

GitHub stars don't tell you what's production-ready. Benchmark posts don't tell you what scales under real load. The volume makes it impossible to stay current by reading changelogs, so engineers default to whatever was popular six months ago, or whatever's already in the codebase, even when better options exist and they know it.

The cost of picking wrong usually shows up later. Vendor lock-in, rewrites, performance ceilings that only appear at scale. The decision gets made quickly and paid for slowly.

How tools make the list

Every tool on PyStack cleared four criteria before it was included.

Real-world use. Stars don't count. If teams I trust aren't shipping with it, it doesn't make the cut.

Active maintenance. Recent commits, responsive issues, docs that aren't stale. A good library that stopped moving is a liability.

A clear job to do. When two tools solve the same problem, only the stronger one ships here. Curation means saying no regularly.

Honest scope. No demos pretending to be frameworks. No tools that promise more than the code actually delivers.

That last one takes some judgment. It's easier to enforce after you've been burned by a tool that looked solid in the docs and turned out to have sharp edges everywhere else.

Where this comes from

My background is in regulated systems. Payments, open banking, the kind of software where a wrong assumption costs more than a bad sprint. Spending years in that environment builds a specific kind of skepticism toward anything that only works cleanly in demos.

PyStack grew directly out of evaluating libraries against real problems, not benchmarks. When you're building something that has to hold up, your opinions about tools get formed fast.

What kind of document this is

It's updated continuously. Tools that fall behind on maintenance get demoted. New ones that meet the bar get added. The methodology is transparent so you can follow the reasoning and disagree with it.

Disagreement is welcome. PyStack is more useful when people push back on the picks. If something's missing, or a choice looks wrong, or the categorization doesn't match how you think about the stack: there's a contact link at the bottom of every page.

The goal isn't to be comprehensive. It's to give you a defensible starting point, built by someone who had to care about getting it right.