DepLab - Python dependency risk advisor

Inspiration

pip's resolver catches only about half of real dependency conflicts. The rest install successfully and crash later at import or runtime. Every Python developer knows the hit-and-trial pain of finding versions that actually work together.

What it does

DepLab is an evidence-aware dependency advisor. Upload requirements.txt, ask an upgrade question like "Can I upgrade numpy to 2.0.2?" and DepLab tells you what will break, why, and the exact version set that works — before you install anything. It strictly separates published-constraint facts (BLOCKED) from ML predictions (risk %).

How we built it

We installed and tested 4,109 real package combinations in isolated environments (uv, Linux x86-64, Python 3.10-3.12), recording where each failed. On this data we trained the DepLab Hybrid model: weighted logistic regression over structured features (wheel/ABI tags, constraints, release timing, changelog signals) plus a post-install failure warning head. On 840 held-out rows from unseen package families: 78.8% accuracy, 80.4% balanced accuracy, 90.4% failure precision, 72.4% failure recall. GPT-5.6 parses questions and explains evidence. The entire pipeline was built with Codex.

Challenges

Avoiding label leakage (test outcomes are labels only, never model inputs), honest evaluation on unseen package families, and keeping facts separated from predictions in every answer.

Try it live

https://deplab.13-234-114-139.sslip.io Demo: https://youtu.be/91f61GPDs7s

Built With

• codex
• fastapi
• gpt-5.6
• machine-learning
• python
• react
• scikit-learn
• uv