​
Data Management

An AIUS project is a plain directory on your machine. You own it — every input, intermediate dataset, and output is a regular file you can open, edit, and version.

​
Project layout

my-project/
  context/                 # your brief: goals, domain notes, definitions (markdown, docs)
  data/
    raw/                   # your source datasets (AIUS moves your files here on first run)
    processed/             # cleaned, analysis-ready datasets AIUS produces
  output/
    discovery/             # data profiling: structure, distributions, anomalies
    notebooks/<analysis>/  # one Jupyter notebook per analysis
    validation/            # model validation folds and predictions
  .aius/                   # project state: current stage, goals.json
Start a project by putting your brief in context/ and your dataset in data/, then run aius in that directory.

​
Supported data formats

CSV, TSV, Excel (.xlsx), Parquet, and JSON. Your data stays on your machine and is read by code running locally. Notebooks execute in a curated, fixed Python environment (pandas, numpy, scikit-learn, XGBoost, LightGBM, matplotlib, seaborn, and more) — you can’t install extra packages into it, which is what keeps runs reproducible.

​
Common how-tos

​
Add or replace a dataset

Drop the new file into data/raw/ (or overwrite the existing one), then tell the agent in chat — e.g. “I’ve added q3_sales.parquet to data/raw, include it in the analysis.” It will profile the new data and fold it in.

​
Change the brief mid-project

Edit your files in context/ (or context/CONTEXT.md, the agent’s written understanding of your brief), then tell the agent what changed. For a significant change of direction, say it explicitly in chat — the agent adapts its goals and plan from the conversation.

​
Correct the agent’s understanding

The pipeline pauses at three review points — use them:
  • Context review — read context/CONTEXT.md. If the agent misread your brief, say so in chat and it rewrites its understanding before doing anything else.
  • Discovery review — check output/discovery/. Wrong column interpretation, missed anomaly? Point it out now, before cleaning.
  • Goal review — the hard gate. The parsed goals live in .aius/goals.json; nothing proceeds until you approve them. This is the cheapest place to fix a wrong assumption.

​
Re-run or tweak an analysis

Every analysis is a real notebook under output/notebooks/<analysis>/notebook.ipynb. Ask the agent to revise and re-run it (“redo the churn model without the leaky feature”), or edit the notebook yourself and ask the agent to continue from it.

​
Regenerate a deliverable

Ask the agent to redo the underlying analysis and republish — deliverables (report, model, deck, brief) are produced from the notebooks, so fixing the notebook and asking for a republish refreshes them on your dashboard.

​
Pick up where you left off

aius --continue        # resume the last session on this project
aius --session <id>    # resume a specific session
Project state (stage, goals, outputs) lives in the project directory, so closing the terminal loses nothing.

​
Start over

aius --reset           # wipe local project state (asks for confirmation)
This resets the pipeline to the beginning. Your files in context/ and data/ are untouched; your API key is preserved.

​
Good to know

  • data/raw/ and data/processed/ are kept out of git by default (AIUS writes the ignore rules) — datasets don’t end up in your repo history.
  • Anything that touches your machine — shell commands, file writes, access outside the project — is gated by a permission prompt before it runs.