.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Auditor-AI
Auditor-AI is a pre-flight checklist for machine learning projects. Instead of waiting for training to fail, a reviewer to miss a subtle bug, or deployment to expose a data problem, a user uploads a zipped ML repo and gets back a fast, structured engineering audit. The FastAPI backend extracts and inventories the repo, then a dataset agent profiles it for missing values, likely target column, class imbalance, and outliers, while a training agent statically scans the code for random seeds, validation splits, optimizer/learning-rate/epoch configuration, and framework signals like PyTorch, sklearn, or XGBoost. A reviewer agent cross-validates both sets of findings against each other - flagging things like missingness with no null-handling logic, or imbalance with no class weighting - and computes a deterministic ML Health Score starting at 100, mapped to an A–F grade. Scoring is entirely rule-based, with no LLM involved, so results stay predictable and explainable. A separate solution agent then calls Groq's llama-3.3-70b-versatile to draft concrete, memory-conscious code fixes based on those findings, packaged into a downloadable suggested-changes.md. The React/Vite frontend handles upload and polling, while Supabase tracks each job's status end to end. Built in hackathon time, Auditor-AI is honest about its scope - static analysis only, first valid dataset and script per repo - but it proves the core idea end to end: turning silent, easy-to-miss ML risks into a fast, legible, and actionable review.
10 Aug 2026
