unloop

Turn slow pandas loops into fast, verified vectorized code.

unloop finds slow patterns in your pandas and NumPy code, like iterrows() loops and apply(axis=1), and rewrites them into fast vectorized operations. Every rewrite is run and compared against your original code before it's offered, so unloop only ever makes your code faster, never different.

%%unloop
for i, row in df.iterrows():
    df.at[i, "total"] = row["price"] * row["qty"]

unloop: 2016x faster via rules, results identical. Running:
df['total'] = df['price'] * df['qty']

Why unloop?

Row-by-row loops are one of the most common reasons pandas code is slow. Rewriting them by hand is easy to get subtly wrong, especially with missing values and data types. unloop does the rewrite for you and proves it's correct on your actual data.

• Verified, not guessed. Both versions run on private copies of your data, plus extra test data with missing values mixed in. If the results differ in any way, the rewrite is rejected.
• Only when it's actually faster. Every rewrite is benchmarked, and unloop only changes your code if it's at least 1.5x faster.
• Your data stays untouched. Verification and timing run on copies, never on your real variables.
• Keeps your formatting. Only the slow lines change; your comments and blank lines stay exactly as you wrote them.
• Private by design. The optional AI runs on your own computer through Ollama. Nothing is sent to the cloud.

Installation

pip install unloop

Requires Python 3.9+ and pandas 2.0+.

Optional: for faster text operations, also install PyArrow:

pip install pyarrow

Quick start

In Jupyter notebooks

Load the extension once per notebook:

%load_ext unloop

Then put %%unloop at the top of any cell:

%%unloop
df["total"] = df.apply(lambda r: r["price"] * r["qty"], axis=1)

What happens next depends on what unloop finds:

To skip the AI and use only the built-in rules:

%%unloop --no-ai

In Python code

import unloop

result = unloop.optimize(code, {"df": df})

print(result["status"])   # "optimized", "no_change", "rejected", or "not_faster"
print(result["code"])     # the optimized code, or your original if not optimized

What unloop can rewrite

Code that doesn't match these rules can be handled by the optional AI fallback.

Optional: AI fallback with Ollama

For slow code that none of the built-in rules recognize, unloop can ask a small coding model running locally on your computer.

1. Install Ollama from ollama.com.

2. Download the model:

   ollama pull qwen2.5-coder:3b
   

That's it. unloop uses it automatically when Ollama is running. AI suggestions go through exactly the same verification and benchmarking as the built-in rules, and if a suggestion fails, unloop tells the AI what went wrong and lets it try again (up to 3 times).

Verified AI rewrites are saved in ~/.unloop/cache.json on your computer, so the same code never needs the AI twice. Saved rewrites are re-verified on your current data every time they're used.

How it stays safe

For every rewrite, unloop:

1. Runs your original code and the new code on separate private copies of your data.
2. Compares every resulting DataFrame: values, columns, index, and data types.
3. Repeats the comparison on a small copy of your data with missing values (NaN) inserted, to catch bugs your current data might not reveal.
4. Benchmarks both versions and keeps the new code only if it's meaningfully faster.

If anything fails, your original code is used.

Limitations

• unloop compares DataFrames. Changes to other kinds of variables (lists, dictionaries, files written to disk) are not checked.
• Verification runs both versions of your code. Avoid using %%unloop on cells with side effects like writing files, sending emails, or calling external APIs.
• AI-generated code is run on copies of your data during verification. It comes from a model on your own machine, but as with any generated code, review it before relying on it in important work.
• The verification is thorough but not a mathematical proof. It tests your data and common edge cases, not every possible input.

License

MIT License. Copyright (c) 2026 Sankalp Gaur. See LICENSE for details.