Skill UI

This skill provides a multi-round, deep critical review of research proposals, papers, or experimental results by simulating the perspective of a senior ML reviewer (e.g., NeurIPS/ICML level). It guides the user through identifying logical gaps, missing experiments, and narrative weaknesses, ensuring high fidelity and iterative refinement of technical work.

This skill automates the end-to-end workflow for deploying and executing ML training and inference experiments. It intelligently detects the target environment (Local GPU, Remote SSH, Vast.ai, Modal serverless) and handles necessary steps like GPU pre-checks, code synchronization (rsync/git), integrating ML tracking tools (W&B), and launching the compute job via screen sessions or serverless deployment. Use when initiating a complex training job.

A comprehensive tool designed to systematically plan ablation studies for machine learning research papers. It guides the process by having an AI (Codex) simulate a rigorous reviewer, identifying necessary experiments to isolate component contributions and answer critical research questions. It also includes a feasibility review step to manage compute resources and implementation complexity.

A comprehensive skill for managing, provisioning, and destroying GPU instances on vast.ai. It analyzes your deep learning task to determine precise hardware requirements (VRAM, GPU count, storage) and searches for cost-optimized, high-reliability GPU offers. The skill handles the entire lifecycle from initial analysis to experiment monitoring, ensuring you only rent what you need at the best possible value.

This skill guides users on setting up and running LLM-as-judge evaluations using Arize. It covers defining evaluators (including prompts, classification choices, and models) and executing tasks against real data spans or experiments. It supports key metrics like hallucination, faithfulness, and correctness, and facilitates continuous monitoring for comprehensive LLM performance analysis.

PufferLib is a high-performance, scalable RL framework designed for massive parallel training and vectorized simulation. Featuring PuffeRL (optimized PPO+LSTM) and the PufferEnv API, it supports multi-agent systems, custom environment creation, and achieves industry-leading speeds (up to millions of steps/second). Ideal for advanced research and large-scale AI experimentation.

A comprehensive framework designed to structure and automate PyTorch deep learning workflows. It simplifies boilerplate code by organizing models into LightningModules, automating training loops via the Trainer, and providing robust support for multi-GPU/TPU scaling (DDP, FSDP, DeepSpeed), data handling (DataModules), and best-in-class experiment logging (W&B, MLflow). Ideal for professional, scalable deep learning project development.

Stable Baselines3 is a robust, PyTorch-based library providing reliable implementations of state-of-the-art reinforcement learning algorithms (PPO, SAC, DQN, etc.). It is designed for standard RL experiments, quick prototyping, and developing complex single-agent systems. The library supports training agents, creating custom Gymnasium environments, and running vectorized simulations for accelerated learning.

This powerful compiler ingests heterogeneous research inputs, including PDF papers, code repositories, experiment logs, and raw notes. It transforms this diverse data into a comprehensive, structured Agent-Native Research Artifact (ARA). The system extracts core knowledge atoms, builds cognitive maps (claims, concepts, heuristics), and generates a verifiable, executable knowledge package for advanced AI consumption and research iteration.

This skill implements Autonomous Optimization (AO) using Hypothesis Tree Refinement (HTR). It iteratively improves concrete artifacts (like code, models, or data pipelines) by running structured, long-term experiments against a measurable objective. It maintains a persistent hypothesis tree, coordinating subagents to test hypotheses in isolation while using a held-out test evaluator to prevent overfitting and ensure reliable, cumulative research progress. Ideal for complex benchmark tuning, agent refinement, or optimizing entire workflows.