What problem does it solve? Choosing the right model, architecture, and deployment strategy in the fast-moving AI/ML landscape is error-prone: benchmarks are contaminated, model rankings shift weekly, and teams overpay for frontier APIs or fine-tune when RAG would suffice. This Skill provides a practitioner reference covering the 2025-2026 landscape so architecture decisions are grounded in current evidence. ## Core Features & Use Cases - Model Selection & Routing: Decision rules for hosted APIs vs open weights, staged model routing (RouteLLM-style), and abstraction layers (LiteLLM/OpenRouter) to avoid lock-in. - Build Pattern Guidance: The prompting → RAG → fine-tuning progression, with production RAG defaults (hybrid BM25+dense retrieval, reranking, contextual retrieval) and LoRA/QLoRA/DPO/GRPO fine-tuning criteria. - Inference & Serving: Quantization format trade-offs (AWQ/GPTQ/GGUF/FP8), serving engine comparison (vLLM/SGLang/TensorRT-LLM), and continuous batching patterns. - Governance & Evaluation: EU AI Act compliance deadlines, benchmark contamination skepticism, and internal eval construction. - Use Case: When asked whether to self-host DeepSeek-V4 or use a hosted API for a high-QPS document pipeline, this Skill provides the cost, privacy, and capability trade-off framework plus the vLLM/SGLang serving configuration to act on the decision. ## Quick Start Ask the AI to recommend a model and deployment strategy for your workload, for example whether to use a hosted frontier API or self-host open weights on vLLM for a RAG application.
