ML workflow auditor

PipelineGuardian

Finds ML bugs your linter can't see — data leakage, reproducibility gaps, and validation-strategy mismatches. Every issue includes supporting evidence.

AST where rules are enough. LLM where context matters.

Data leakage Reproducibility Validation strategy review

How it works

1

Static analysis

Parses your code's structure directly and checks it against known leakage and reproducibility patterns.

deterministic
2

Schema inspection

Looks at your actual dataset for timestamp columns, grouped entities, and class imbalance.

deterministic
3

Contextual review

An LLM weighs in only when a schema signal makes the split strategy worth reasoning about.

conditional

Example output

high deterministic scaler_fit_before_split

Scaler is fit on the full dataset before train_test_split — statistics from the test set leak into training.

Evidence

scaler.fit_transform(X)

Try an example — loads a real file from the repo, no upload needed

Drop notebook or script here, or click to browse
.ipynb or .py
Drop a CSV here to enable schema checks
Timestamp / group / imbalance detection
Advanced (local dev)

Only visible on localhost — production always targets the deployed backend.