Home / Guides / Guide

10 GitHub Repos Every AI Engineer Should Know (By Job Postings)

By Saurav Sharma · November 18, 2026

Job postings name tools. Interviewers ask about the tools their team runs. This list takes ten open-source repos and ranks them by how many companies name them in AI-engineering job postings, so you can pick where to spend a weekend. Each entry says what to read, what to build, and which interview questions it helps with.

The ten repos, ranked

Counts are from aijobprep's study of 3,647 AI-engineering job postings at 198 companies, collected August 2026. A company counts once if any of its job titles or descriptions names the tool. Most searches ignore capitalization. Words with other meanings got stricter rules: JAX had to be uppercase, and Ray had to appear next to another tool name. All ten repos had new commits in the week before 2026-11-18, and none is archived.

#RepoCompanies naming the toolStars (2026-11-18)
1kubernetes/kubernetes121129,256
2pytorch/pytorch107105,688
3tensorflow/tensorflow74201,611
4modelcontextprotocol/python-sdk5824,896
5jax-ml/jax4936,649
6langchain-ai/langchain48149,869
7ray-project/ray4544,321
8vllm-project/vllm3997,174
9langchain-ai/langgraph3045,121
10huggingface/transformers29169,233

Stars and push dates come from the GitHub API on 2026-11-18. All ten repos were pushed within the last two days and none is archived.

How to read this ranking

The repos

1. kubernetes/kubernetes

121 companies. Kubernetes schedules and runs containers across a cluster. Teams use it to run training jobs and keep model servers available when a container fails.

Build in a weekend: Read pkg/scheduler and pkg/controller/deployment. Then write a Deployment for a model server that requests a GPU. Break it on purpose, for example with an image name that does not exist, and read the pod events to find the message that explains the failure.

Helps with: "How do you deploy and scale a model server?", "A pod keeps restarting. Walk me through it.", GPU scheduling and autoscaling questions for MLOps roles.

2. pytorch/pytorch

107 companies. PyTorch is a Python library for building and training neural networks.

Build in a weekend: Read torch/nn/modules/linear.py and the torch/autograd folder. Then write a training loop by hand with no Trainer class: a model, a loss, a backward pass, an optimizer step. Add mixed precision (running some math at lower precision to save memory), then save a checkpoint, stop training, and resume from it.

Helps with: "Explain what loss.backward() does.", "Why does the loss become NaN (not a number)?", "How would you fit a bigger model on one GPU?"

3. tensorflow/tensorflow

74 companies. TensorFlow is Google's machine learning framework. It is older than PyTorch and still named in postings, often next to PyTorch in a "TensorFlow or PyTorch" line.

Build in a weekend: Skim tensorflow/python/keras. Export one small model to SavedModel format, load it in a second script, and compare its output with the original. You only need enough to say how it differs from PyTorch.

Helps with: "PyTorch or TensorFlow, and when does it matter?", "How do you move a trained model into production?"

4. modelcontextprotocol/python-sdk

58 companies. The Model Context Protocol (MCP) is a standard way for an AI app to call outside tools and data. This is the official Python SDK (software development kit). The counting rule for this row also matches the short form "MCP", so it includes postings that never spell out the full name.

Build in a weekend: Read src/mcp/server and the examples folder. Build a server with two tools that read a local folder, connect it to an AI client, and write down what happens when a tool returns an error.

Helps with: "How do agents call tools safely?", "How would you give a model access to an internal system?", prompt injection and permission questions.

5. jax-ml/jax

49 companies. JAX is a numerical library that computes gradients, compiles, and batches Python and NumPy code for GPUs and TPUs (Google's Tensor Processing Units). Several large model labs use it for training. Matched case-sensitively as "JAX" so first names do not count.

Build in a weekend: Read jax/_src/pjit.py and run the examples. Write a tiny model with jit, grad, and vmap, then time it with and without compilation.

Helps with: "What does JIT (just-in-time) compilation buy you?", "How do you shard a model across devices?", numerical performance questions for research engineering roles.

6. langchain-ai/langchain

48 companies. LangChain is a library of building blocks for apps that call language models: prompts, tools, retrievers, output parsers.

Build in a weekend: Read libs/core. Build a retrieval question-answering script twice: once with LangChain, once with plain API calls. Compare the line counts, then list what the library gave you and what it hid.

Helps with: "Build a RAG (retrieval-augmented generation) app, which finds relevant documents and gives them to the model to answer a question, and explain each step.", "When would you skip a framework?", "How do you debug a chain that returns a wrong answer?"

7. ray-project/ray

45 companies. Ray runs Python code across many machines. Its libraries cover data loading, training, tuning, and model serving. Matched case-sensitively as "Ray" next to a tool name, because the word is also a first name.

Build in a weekend: Read python/ray/serve and python/ray/data. Run a local Ray cluster, serve a small model with Ray Serve, then send 100 parallel requests and watch where the time goes.

Helps with: "Design batch inference for 10 million documents.", "How do you scale a serving layer?", distributed training and retry questions.

8. vllm-project/vllm

39 companies. vLLM is an engine that serves large language models with high throughput. Its key ideas are paged attention (storing the attention cache in small blocks) and continuous batching (adding new requests mid-run).

Build in a weekend: Read the scheduler in vllm/v1/core/sched. Serve a small open model, send more and more requests at the same time, and plot how long each request takes against how many requests finish per second.

Helps with: "Why is the first token slow and the rest fast?", "How do you cut inference cost?", GPU memory and batching questions.

9. langchain-ai/langgraph

30 companies. LangGraph models an agent as a graph of steps that share state, with support for saving progress and pausing for human review.

Build in a weekend: Read libs/langgraph/langgraph/pregel. Build a three-step agent (plan, call a tool, check the result) with a checkpoint, kill it midway, and resume it.

Helps with: "How do you stop an agent from looping forever?", "How do you add human approval?", state and failure-recovery questions.

10. huggingface/transformers

29 companies. Transformers provides code to load and run pretrained open models. Its Trainer class helps you fine-tune a model on your own examples. The count is companies whose postings name Hugging Face as a tool. Mentions where Hugging Face was an employer or a customer were dropped.

Build in a weekend: Read src/transformers/models/llama/modeling_llama.py from top to bottom, then trainer.py. Fine-tune a small model on a few hundred examples and compare its answers before and after.

Helps with: "Explain attention.", "How do you fine-tune with limited data?", "When is fine-tuning the wrong answer?"

How to use a weekend on one repo

  1. Saturday morning: read the folder named above. Write five sentences on what the code does and why it is built that way.
  2. Saturday afternoon: run the repo's own example. Change one parameter and record what changed.
  3. Sunday: build the small project in the entry. Keep a short README with the commands, one result, and one thing that failed.
  4. Sunday evening: practice explaining the project aloud in two minutes: what you read, what you built, what broke, what you would change. That story answers the "tell me about a project" question.

Pairs that make a stronger story

Other tools in the same study

Three tools had counts close to or above the bottom of the list. Apache Airflow (42 companies) schedules data pipelines and is not specific to AI work, so it is not in the ten. SGLang (30 companies) serves language models like vLLM does, and one serving engine is enough for a weekend. MLflow (21 companies) tracks experiments and model versions, and is worth a look for MLOps roles.

What to check before you start

Next step

Once you have a repo story, test it against real questions. The AI Engineer question bank covers RAG, agents, evals, and cost. Related banks: ML Engineer, MLOps Engineer, and Forward Deployed Engineer.