Jupyter Notebook Coach-AI Jupyter coding assistant

Example

Given a snippet that fails or produces incorrect output, the Coach explains the error, offers minimal reproducible test cases, and provides corrected code and alternatives. Example corrections include fixing common pandas mistakes (chained assignment), resolving scikit-learn API changes when loading/saving models, and recommending explicit dtype conversions to avoid silent bugs.

Scenario

You load a CSV and get wrong aggregation results. The Coach suggests checking for hidden whitespace and mixed types, proposes `df.columns = df.columns.str.strip()` and `df['col'] = pd.to_numeric(df['col'], errors='coerce')`, and supplies a short test cell to validate results. It also shows how to trap runtime exceptions with `%pdb` and how to use `import traceback; traceback.print_exc()` for richer tracebacks inside notebooks.

Example

Advice and templates for structuring notebooks (title + description cell, inputs/outputs, sections for data loading, preprocessing, modeling, evaluation), plus concrete commands to produce reproducible artifacts. Example workflow: capture environment with `pip freeze > requirements.txt` or `conda env export -n myenv > environment.yml`, parameterize a notebook with papermill, and produce a shareable HTML via `jupyter nbconvert --to html analysis.ipynb`.

Scenario

A research team must reproduce an experiment. The Coach provides a checklist: pin package versions, add a metadata cell with dataset version and random seeds, use `%%bash` cells or Dockerfile snippets to show how to create a reproducible environment, and demonstrates using `papermill` to run the notebook with different parameter sets for batch experiments. It also shows how to convert the notebook to a Python module with `nbconvert --to script` and how to break reusable functions into a package for unit testing.

Example

Shows how to profile slow cells (`%timeit`, `%lprun` from line_profiler, `%%prun`), then provide alternative implementations: vectorized pandas ops, use of categorical dtypes, chunked processing with Dask or pandas.read_csv(..., chunksize=...), JIT acceleration with numba, or parallelism via joblib. Sample transformation: replace a slow row-wise apply with a vectorized expression, e.g. `df['score'] = df['x'] * 0.3 + df['y'] * 0.7` instead of `df.apply(lambda r: r['x']*0.3 + r['y']*0.7, axis=1)`.

Scenario

An analyst processes a 20GB CSV in a notebook and steps through performance improvements: first sample subset to reproduce the issue, profile to find hotspot, convert costly row-wise operations to vectorized code, and finally recommend Dask or out-of-core processing if memory remains a bottleneck. The Coach includes example code for chunked aggregation: total = 0 for chunk in pd.read_csv('big.csv', chunksize=1_000_000): total += chunk['value'].sum() and a Dask alternative using `dask.dataframe` for a nearly drop-in replacement.