Back to Products
Open Source Python Library / Developer Tool2026 (v0.2.0 Released)
FeatureSmith
Open-Source Python Feature Engineering Toolkit & Dataset Review Platform
The Problem
Tabular datasets used in production machine learning models often suffer from unmonitored target leakage, silent schema drift, missingness spikes, and distribution anomalies. Unlike source code which undergoes strict automated linting and code reviews, datasets are often fed into compute-heavy model training loops blindly.
Why I Built It
While building machine learning models for demand forecasting and reinforcement learning, I repeatedly observed that dataset bugs—such as future timestamps leaking into training features or identifier columns correlating with target labels—caused massive delays. I built FeatureSmith to create a zero-friction, automated dataset reviewer that halts invalid pipelines before compute is wasted.
The Solution
FeatureSmith provides a high-performance Python SDK and CLI powered by Polars that executes 8 automated dataset reviewers, generates explainable 0–100 ML readiness scorecards, flags 6 named target leakage patterns, and provides snapshot version diffing (fs.diff()) with a deterministic exit-code CI/CD gate.
Architecture & Pipeline Design
SDK & CLI Interface: Polars-backed dataset parser with low memory overhead
Review Engine: 8 automated reviewers checking schema, missingness, duplicates, cardinality, and distribution bounds
Leakage Detection Suite: 6 specialized detectors flagging correlations, timestamp drift, and ID shape leakage
Scorecard & Report Generator: Explainable 0-100 ML Readiness score with fix recommendations
CI/CD Exit Gate: Deterministic exit-code (0 = clean, 1 = findings) for automated deployment pipelines
Key Engineering & Product Challenges
Handling massive tabular datasets efficiently without exhausting memory during distribution profiling
Formulating explainable, mathematical scoring weights for the 0–100 ML Readiness Scorecard
Detecting subtle target leakage patterns without triggering false positives on legitimate predictor features
Key Trade-Offs & Design Decisions
Selected Polars over Pandas for blazing-fast columnar execution speeds and lower memory overhead
Engineered strict MyPy type hints (100% type safety) and Ruff linting for professional library maintenance
Implemented deterministic exit codes to enable seamless integration into GitHub Actions CI workflows
Technologies & Tools
PythonPolarsCLISDKPyPIMyPyRuffGitHub ActionsApache 2.0
Future Roadmap & Next Version
v0.3.0 release with real-time streaming dataset validation adapters
Automated PyArrow integration for ultra-large Parquet file reviews
Exportable HTML and SVG interactive audit report generators