​

Features

Fast is easy to fake. Reproducible isn’t.

Built for people who have to defend every number to a stakeholder, a regulator, or a reviewer.

​

The pipeline

AIUS works through a structured pipeline on your project, with review gates where you approve before it continues:

1. Context — reads your brief from context/ and writes back its understanding (CONTEXT.md). You review it before anything else happens.
2. Discovery — profiles your raw data: structure, distributions, anomalies. The findings land in output/discovery/ for you to inspect.
3. Goals — turns the brief into explicit, measurable goals with success criteria. A hard gate: nothing proceeds until you approve them.
4. Cleaning — produces processed, analysis-ready datasets in data/processed/.
5. Analysis — pursues each approved goal in real Jupyter notebooks: feature work, modeling, validation, visualization.
6. Report — packages the results into deliverables published to your dashboard.

You can chat with the agent at any point to steer, correct, or dig deeper.

​

Review gates & permissions

• Three review points — your brief interpretation, the data discovery, and the goal list each pause for your input. Goals are a hard gate.
• Permission prompts — anything that touches your machine (shell commands, file writes, access outside the project) asks first. Approve once, always, or reject.

​

Real notebooks, reproducible by default

• Every analysis is a Jupyter notebook under output/notebooks/ — not a hidden scratchpad.
• Notebooks run in a curated, fixed Python environment (pandas, numpy, scikit-learn, XGBoost, LightGBM, matplotlib, seaborn, and more) — consistent results across runs and machines.
• Re-run or edit any notebook yourself; the work is yours.

​

Deliverables on your dashboard

Each run can publish four deliverable types to your dashboard:

• Report — the full findings with charts
• Model — the trained model and its validation
• Deck — a presentation-ready summary
• Brief — the executive short form

​

Sessions you can come back to

• aius --continue resumes your last session; aius --session <id> picks a specific one.
• Project state lives with the project — stages, goals, and outputs persist between launches.
• Every run is recorded end to end (stages, model calls, tool calls, artifacts), so there is a single source of truth for how each insight was produced.

​

Model choice

Pick the model for a run with aius --model <provider/model>. Model traffic is routed through the AIUS service — you never manage provider keys yourself.

​

Works with your data formats

CSV, TSV, Excel, Parquet, and JSON — dropped into your project’s data/ directory. See Data Management for the common workflows.