Skill UI

This comprehensive workflow guides the end-to-end implementation of Retrieval-Augmented Generation (RAG) systems. It covers critical stages, including selecting optimal embedding models, setting up vector databases, designing chunking strategies, implementing hybrid retrieval, integrating Large Language Models (LLMs), and performing rigorous evaluation. Use this when building knowledge-grounded AI applications or advanced semantic search systems.

This comprehensive guide demonstrates how to build and test a voice agent using the Retell AI SDK in TypeScript. It walks through configuring the underlying LLM with a custom persona, creating the voice agent instance, and finally initiating both outbound phone calls and web calls through the API. This is essential for integrating AI-powered voice communication into customer service platforms, call centers, or telephony automation systems.

This resource provides production-ready, structured patterns for building robust voice agent applications using the Retell AI SDK. It covers best practices such as implementing singleton client management, defining typed agent configurations with integrated LLMs, adding automatic retry logic for call failures, and managing large-scale outbound campaigns. Use this when building stable, scalable, and reliable voice automation systems.

A comprehensive, model-agnostic toolkit for advanced LLM development. It enables iterative prompt optimization through static analysis (token usage, clarity), measures critical RAG metrics (relevance, faithfulness, coverage), and validates complex agent workflows and configurations. Essential for rigorous LLM quality assurance and cost management.

SGLang is a high-performance serving framework for LLMs and VLMs that uses RadixAttention prefix caching to deliver structured outputs (JSON/regex/grammar) and agentic workflows with function calls, providing up to 5× faster inference than vLLM across multi-GPU production deployments.

SimPO (Simple Preference Optimization) is a state-of-the-art, reference-free method designed for aligning Large Language Models (LLMs) using human preference data. It is an efficient alternative to DPO and PPO, notably outperforming DPO without requiring a separate reference model. It is ideal for practitioners who need faster, simpler, and more resource-efficient fine-tuning for preference alignment.

Guides post-training of GLM-class LLMs with Megatron-LM and SGLang, focusing on GRPO rollout generation, custom data buffers, and tight RL scaling workflows for research and production use.

Speeds up LLM inference by combining speculative decoding with Medusa multiple heads and lookahead techniques so chatbots, code assistants, and other real-time workloads can run with 1.5-3.6× lower latency while preserving target-model quality.

Optimizes large language model inference on NVIDIA GPUs with TensorRT, delivering 10-100× faster throughput, quantized precision (FP8/INT4), multi-GPU scaling, and serving-ready tooling for production deployments.

This skill provides advanced guidelines for designing robust and effective tools specifically for autonomous AI agents. It shifts the focus from traditional developer APIs to contracts optimized for Language Models (LLMs). Key topics covered include treating tool descriptions as prompt engineering, implementing namespacing, applying the Consolidation Principle, and utilizing Architectural Reduction patterns (like file system access) to maximize agent reliability and workflow complexity.

Provides production-ready Megatron-Core workflows for training 2B–462B parameter LLMs with tensor/pipeline/expert/context parallelism, maximizing GPU efficiency on H100 clusters.

Guides training large language models with Volcano Engine's verl library, covering PPO/GRPO workflows, distributed backends, multi-turn rollouts, reward configs, critic-enabled PPO, and production deployments using Ray workers.