A cookbook for the "just glue it together" era of AI

What it does This repo is a curated grab-bag of practical AI projects — OCR pipelines, RAG systems, multi-agent workflows, and multimodal apps — each designed as a standalone reference you can crib from or adapt. Think of it as a field guide for the current moment where most “AI engineering” means picking the right model, the right framework, and the right API, then making them talk to each other without catching fire.

The interesting bit The breadth is the point. One folder has a GitHub PR review agent wired to Telegram via MCP; another has a prescription digitizer that validates drug names against RxNorm in real time. The author isn’t chasing novelty per project — they’re mapping the combinatorial space of what actually gets built right now, from a “rock music RAG” using BM25 to a Cal.com scheduling agent. It’s a snapshot of the tooling landscape as much as the use cases.

Key highlights

• 25+ projects across agents, OCR, audio, multimodal, and RAG, each with its own folder and README
• Heavy use of current frameworks: LangGraph, Agno, smolagents, AG2, plus MCP servers for GitHub, Trivago, and Cal.com
• Model promiscuity by design — jumps between OpenAI, Anthropic, Google Gemini, Mistral, DeepSeek, MiniMax, Qwen, and local Ollama models depending on the task
• Several projects mix retrieval architectures: GraphRAG with NetworkX, hybrid vector-plus-graph, HyDE with hypothetical embeddings, and database-routed RAG with fallback web search
• Contributing guidelines enforce structure: requirements.txt, .env.example, snake_case naming, one project per PR

Caveats

• “Production-ready” is claimed in the top-level description, but the READMEs themselves are the only source — no tests, CI status, or deployment evidence visible
• Every project appears to need its own API keys and setup; the cumulative keyring for the full repo would be substantial
• Some project descriptions read like framework marketing copy (“powered by MiniMax M2.7”), which doesn’t clarify what’s novel versus configured

Verdict Worth bookmarking if you’re a developer who learns by disassembling working examples, or if you need to quickly sanity-check how others are wiring together RAG, agents, and MCP in 2024. Skip it if you want deep theoretical treatment or a unified framework — this is explicitly a scattered collection, not a platform.