Data Science & ML Claude 스킬

LLM Evaluation with Evidently.ai

Evaluates LLM output quality and optimizes prompt templates using Evidently.ai metrics and LLM-as-a-Judge patterns.

Reward Model Training

Streamlines the development and training of reward models for RLHF pipelines and thinking quality scoring.

Context Optimization

Optimizes AI agent context through compression, masking, and strategic partitioning to maximize token efficiency and model performance.

LangChain Framework

Builds sophisticated LLM applications using LangChain for prompt management, model chaining, and structured output parsing.

LLM Project Development Methodology

Architects and optimizes LLM-powered applications using structured methodologies, pipeline design, and agent-assisted development patterns.

Model Quantization

Optimizes large language models for efficient inference and training by reducing memory footprint using advanced precision-shifting techniques like 4-bit and 8-bit quantization.

Semantik Plugin Development

Builds custom plugins for the Semantik search engine to handle document ingestion, embeddings, and AI-powered reranking.

Auto-Claude Memory System

Implements a persistent knowledge graph memory system for Auto-Claude to retain context and learn patterns across development sessions.

BDI Mental State Modeling

Transforms external RDF context into formal Belief-Desire-Intention (BDI) models to enable rational agency and explainable reasoning in AI agents.

Vision Model Fine-Tuning

Fine-tunes vision-language models like Pixtral and Ministral using Unsloth's FastVisionModel optimizations for faster training.

DPO Preference Optimization

Optimizes Large Language Models using Direct Preference Optimization to align behavior with preferred response pairs without explicit reward modeling.

PEFT & LoRA Fine-Tuning

Streamlines parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and Unsloth to optimize memory and training speed.

Memory Manager

Persists project knowledge and error-solution patterns across autonomous coding sessions to prevent repetitive mistakes.

GRPO Alignment & RLHF

Implements Group Relative Policy Optimization for efficient LLM alignment and reinforcement learning from human feedback.

Context Engineering Fundamentals

Provides foundational expertise in context engineering to optimize AI agent performance and manage token usage effectively.

Ollama Python Integration

Provides a clean, Pythonic interface for local LLM inference, chat completions, and model management using the official Ollama library.

Multi-Agent Architecture Patterns

Designs and implements sophisticated multi-agent systems using supervisor, swarm, and hierarchical patterns to solve complex context management challenges.

xAI Grok Model Selector

Simplifies the selection and optimization of xAI Grok models for development tasks through detailed capability and cost analysis.

HiPerGator Multi-Group Conda Manager

Configures and manages multiple Conda environment locations across different research group storage allocations on the UF HiPerGator supercomputer.

Context Degradation Diagnostics

Diagnoses and mitigates AI agent performance failures caused by long-context attention loss, poisoning, and informational clash.

Book SFT & Style Transfer Pipeline

Builds high-quality fine-tuning datasets from literary works to train AI models in specific authorial voices and writing styles.

Amazon Bedrock Knowledge Bases

Builds and manages fully managed RAG solutions using Amazon Bedrock for semantic search and document-based AI applications.

QLoRA Experimentation & Optimization

Optimizes LLM fine-tuning via advanced QLoRA patterns, hyperparameter tuning, and memory-efficient implementation strategies.

Rust Machine Learning & AI

Guides the development of high-performance ML and AI applications in Rust using memory-efficient patterns and GPU acceleration.

Memory System Architecture

Design and implement multi-layered memory architectures for AI agents to ensure long-term state persistence and entity consistency.

Vector Index Tuning

Optimizes vector database performance by balancing search latency, recall accuracy, and memory footprint.

ComfyUI Stable Diffusion Manager

Manages ComfyUI instances for node-based Stable Diffusion image generation with automated GPU configuration and model management.

FiftyOne Dataset Manager

Manages FiftyOne dataset visualization and curation environments using Podman Quadlet containers with integrated MongoDB sidecars.

Memory Manager

Persists learned solutions and error patterns across autonomous coding sessions to create a self-healing development environment.

Physics Analysis Spec Builder

Generates and refines structured physics analysis specifications using standardized templates and domain-specific best practices.