📘 Awesome AI Engineering

The Full-Stack LLM Engineering Playbook

A full-stack LLM engineering playbook — practical guides for building, deploying, and evaluating LLM systems and AI agents. For a deep dive into the research frontier, explore the 📖 LLM Technology Landscape & Evolution — a curated reading list covering the full LLM stack, from model architectures and training, fine-tuning, inference optimization, reasoning, and Agent systems.

📑 Table of Contents

📚 Content	🔗 Quick Link
Building AI Agents	🔍 Explore
Building LLMs for Production	🔍 Explore
Building High-Performance, Private AI Infrastructure for the Enterprise	🔍 Explore
Mastering the Model Context Protocol (MCP)	🔍 Explore
Agent Memory Part I (A Survey of Memory)	🔍 Explore
Agent Memory Part II (Building Memory Modules for Agentic AI Systems)	🔍 Explore
Agent Evaluation (Eval) Engineering	🔍 Explore

---

📚 Building AI Agents

🔑 Mind Map (Key Concepts)

📥 Download High-Resolution Mind Map (.jpg)

🔍 Click here to unfold the full Mind Map (agents-architecture-operations-and-evolution-mindmap.jpg)
(点击展开完整思维导图)

📑 Presentation Slides

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View the "Building AI Agents" Slides (PDF)
📥 Download PDF (Direct Link)

🛠️ Hands-on: A Minimal ReAct Agent

👉 View the AI Agent Project in the LLMs-Lab repository on the Eric-LLMs GitHub profile.

To bridge theory with practice, I developed a modular AI Agent project that implements autonomous reasoning and task execution:

• Architecture: Utilizes a decoupled structure with dedicated directories for Agent logic, Tools, Utils, and Prompts.
• Reasoning Loop: Features an AutoGPT.py implementation using ReAct (Reasoning and Acting) logic to handle complex, multi-step goal decomposition.
• Functional Tools: Includes custom tools for deep data analysis (Excel processing via Pandas), automated communication via email, PDF-based QA interrogation (FileQATool), requirements-driven document generation (WriterTool), and dynamic script-based auditing of structured files using custom heuristics and thresholds (PythonTool).
• End-to-End Workflow: Supports real-world scenarios, such as identifying underperforming suppliers from sales records and autonomously drafting/sending notifications.

🧰 Key Open-Source Projects & References

The following open-source projects represent prominent examples of agentic AI engineering:

Project	Description	Key Strengths
Claude Code	Anthropic's official terminal-based agentic coding tool	Agentic coding, terminal-native, full codebase understanding, git workflows
OpenAI Codex	OpenAI's open-source agentic coding CLI	Agentic coding, terminal-native, sandboxed execution, bash tool use
Hermes-Agent	Self-improving AI agent with built-in learning loop	Skill creation from experience, cross-session memory, multi-channel (CLI/Telegram/Discord/Slack)
OpenClaw	Personal AI assistant, local-first, any OS/platform	Local-first Gateway, multi-channel messaging, voice support, session & tool management
claurst	Community-maintained reference implementation of Claude Code	Internal architecture study, reverse-engineering insights, codebase structure reference

These projects showcase diverse agent architectures — from developer-focused coding agents (Claude Code/OpenAI Codex/claurst) to general-purpose personal assistants (OpenClaw) and self-learning agents (Hermes-Agent). Studying their design decisions is valuable for building your own agent systems.

⬆️ Back to Top : Table of Contents

---

📚 Building LLMs for Production

This guide covers LLM production, from Transformer architectures to advanced techniques like RAG and Fine-Tuning. It explores frameworks like LangChain, methods to mitigate hallucinations, and optimization via quantization. Learn to build autonomous agents for real-world use.

🔑 Mind Map (Key Concepts)

📥 Download High-Resolution Mind Map (.jpg)

🔍 Click here to unfold the full Mind Map (building-llms-for-production-mindmap.jpg)
(点击展开完整思维导图)

📑 Presentation Slides

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View the "Building LLMs for Production" Slides (PDF)
📥 Download PDF (Direct Link)

🛠️ Hands-on Lab & Examples

👉 Explore Practical LLM Implementations in the LLMs-Lab repository on the Eric-LLMs GitHub profile.

The production-grade principles discussed in this guide—including Fine-Tuning, RAG optimization, LangChain, Prompt Engineering, Function-Calling, Agent, etc.—have each been researched as a standalone module, and each module features multiple project implementations.

⬆️ Back to Top : Table of Contents

---

📚 Building High-Performance, Private AI Infrastructure for the Enterprise

Covers the full-stack AI infrastructure for the enterprise — AI chips, compute clusters, high-speed networking, distributed training, inference serving, cluster scheduling, and secure private deployment.

🛠️ Hands-on Projects & Tools

I. AI Infrastructure

• AIInfra — AI Infrastructure Reference: An open-source reference covering the full-stack AI infrastructure for LLMs — from AI chips, compute clusters, and high-speed networking, to distributed training, inference optimization, and deployment — a hands-on resource for building high-performance, private AI infrastructure for the enterprise.

II. Inference Serving — High-Performance LLM Serving

• vLLM: A high-throughput, memory-efficient LLM serving engine (PagedAttention, continuous batching, prefix caching) — the de facto standard for high-performance private inference.

• SGLang: A fast, structured-generation runtime for LLM inference, complementing vLLM with radix attention and efficient prefix reuse.

III. Training, Pre-training & Fine-tuning

• DeepSpeed: Microsoft's deep learning optimization library — ZeRO memory optimization, mixed precision, and system optimizations for training and fine-tuning models at massive scale.

• Megatron-LM: NVIDIA's large-scale language model training framework — tensor, pipeline, and sequence parallelism, often paired with DeepSpeed for pre-training.

• Unsloth: A fast, memory-efficient fine-tuning library — up to 2x faster and 70% less memory for LoRA/QLoRA fine-tuning of LLMs.

• LLaMA-Factory: A config-driven fine-tuning platform supporting LoRA, QLoRA, and full-parameter tuning across many open LLMs.

• HuggingFace PEFT: The standard parameter-efficient fine-tuning library — LoRA, QLoRA, and more — widely used for enterprise model customization.

IV. Cluster Scheduling & Orchestration

• Volcano: A Kubernetes-native batch system with advanced GPU scheduling and job queueing — widely adopted for enterprise AI clusters.

• KubeRay: A Kubernetes operator for running Ray clusters, bridging distributed compute with cloud-native orchestration.

V. Unified Gateway

• LiteLLM: A unified LLM gateway with an OpenAI-compatible API — model routing, rate limits, budgets, and logging for enterprise private deployments.

VI. Distributed Computing

• Ray: A unified distributed framework for AI training, inference, and serving at scale.

VII. Lightweight Private Deployment

• Ollama: The simplest way to run LLMs locally — a lightweight, self-hosted private deployment option.

• LocalAI: A local, OpenAI-compatible, self-hosted inference server for private model deployment.

VIII. Security & Guardrails

• NeMo Guardrails: NVIDIA's programmable guardrails framework for conversational AI — input, output, and retrieval rails.
• PurpleLlama / Llama Guard: Meta's Llama security toolkit — Llama Guard content-safety classifier and Prompt Guard injection detection.
• Garak: NVIDIA's LLM vulnerability scanner for automated red-teaming.
• LLM Guard: Protect AI's input/output security library for detecting prompt injection and sanitizing LLM traffic.
• Microsoft Presidio: PII detection and data anonymization for compliance in enterprise AI deployments.

⬆️ Back to Top : Table of Contents

---

📚 Mastering the Model Context Protocol (MCP)

A deep dive into the Model Context Protocol (MCP) — the open standard that connects AI agents to tools and data sources. Covers the protocol architecture, official SDKs, and production server implementations.

🔑 Mind Map (Key Concepts)

📥 Download High-Resolution Mind Map (.jpg)

🔍 Click here to unfold the full Mind Map (mastering-the-model-context-protocol-mindmap.jpg)
(点击展开完整思维导图)

📑 Presentation Slides

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View the "Mastering the Model Context Protocol (MCP)" Slides (PDF)
📥 Download PDF (Direct Link)

🧰 Key Frameworks & Tools

The official Model Context Protocol SDKs and reference implementations for building and connecting MCP servers:

• typescript-sdk: The official TypeScript SDK for building MCP servers and clients.
• python-sdk: The official Python SDK for building MCP servers and clients.
• servers — Official Reference Implementations: The official collection of reference MCP servers, including filesystem, fetch, git, memory, and sequential thinking.

🛠️ Hands-on Projects & Tools

👉 Explore Model Context Protocol (MCP) Projects on GitHub A curated collection of industry-standard Model Context Protocol (MCP) server implementations.

⬆️ Back to Top : Table of Contents

---

📚 Agent Memory Part I

A survey of academic research on how agent memory is designed and categorized (forms, functions, dynamics).

🔑 Mind Map (Key Concepts)

📥 Download High-Resolution Mind Map (.jpg)

🔍 Click here to unfold the full Mind Map (unforgettable_agents_architecting_ai_memory-mindmap.jpg)
(点击展开完整思维导图)

📑 Presentation Slides

A Blueprint for Memory in Agentic Intelligence

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View the "A Blueprint for Memory in Agentic Intelligence" Slides (PDF)
📥 Download PDF (Direct Link)

Unforgettable Agents Architecting AI Memory

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View the "Unforgettable Agents Architecting AI Memory" Slides (PDF)
📥 Download PDF (Direct Link)

📑 Further Reading / Resources

For a comprehensive list of papers related to Agent Memory, we highly recommend checking out:
👉 Agent-Memory-Paper-List by Shichun-Liu.

⬆️ Back to Top : Table of Contents

---

📚 Building Memory Modules for Agentic AI Systems

A comprehensive guide on designing memory systems for AI Agents. This document synthesizes academic surveys with practical implementation strategies — covering the taxonomy of agent memory (forms, functions, dynamics), deep dives into Mem0, Letta (MemGPT), and LangMem, and enterprise-grade solutions using Amazon Bedrock AgentCore.

🔑 Mind Map (Key Concepts)

📥 Download High-Resolution Mind Map (mindmap.png)

🔍 Click here to unfold the full Mind Map
(点击展开完整思维导图)

📑 Presentation Slides

Building Memory for Agentic AI: Theory, Frameworks, and Practice

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View Slides (PDF)
📥 Download PDF (Direct Link)

🧰 Key Frameworks & Code Samples

The following frameworks and repositories are discussed in this guide, representing the current state-of-the-art in Agentic Memory:

• Mem0: A dual-layer memory framework supporting working, factual, and semantic memory types for agent state persistence.
• Letta (MemGPT): Manages infinite context by treating agents like an OS with virtual memory and recursive summarization.
• LangMem: A LangChain library that implements Semantic, Episodic, and Procedural memory integration for LangGraph agents.
• Zep / Graphiti: Zep's temporal knowledge-graph framework — builds a dynamic, time-aware memory graph for agent state with causal event support.
• Amazon Bedrock Samples: A comprehensive collection of examples for using Amazon Bedrock, including various implementations of Agentic workflows and memory patterns.

⬆️ Back to Top : Table of Contents

---

📚 Agent Evaluation (Eval) Engineering

Evaluating AI Agents requires a fundamental shift from simple output checks ("vibe checks") to analyzing multi-step trajectories, environment changes, and tool usage. This repository consolidates frameworks and engineering practices for moving from intuition to instrumentation.

---

🔑 Key Considerations

Consideration	Description
The Intuition Trap	Why manual "vibe checks" fail as complexity scales.
The Harness	Building a standardized environment for agent execution composed of Inputs, Tasks, and Graders.
Trajectory vs. Outcome	Evaluating the journey (reasoning logs, tool calls) rather than just the destination (final answer).
Reliability Metrics — Pass@k	Can the agent succeed at least once in k tries? (Good for brainstorming).
Reliability Metrics — Pass^k	Can the agent succeed every single time in k tries? (Critical for autonomous agents).
Swiss Cheese Model	Layering defenses (Automated Evals → Human Review → Production Monitoring) to ensure reliability.
LLM-as-a-Judge	Using LLMs to grade outputs — with known biases (position bias, self-preference) that need calibration.
Task Benchmarks	Standardized suites (SWE-bench, GAIA, AgentBench, τ-bench, WebArena) for measuring real-world task success.
Tool-Call Correctness	Verifying the right tool, right arguments, and right timing — beyond just the final answer.
Process Supervision	Grading intermediate reasoning and tool-call steps, not only the outcome, to catch errors early.
Adversarial Robustness	Stress-testing against prompt injection and goal hijacking.
Agent Security	Testing permission boundaries, tool authorization, and data-handling safety — ensuring the agent cannot overstep access or leak sensitive data.
Cost & Latency	Token efficiency, wall-clock time, and per-task budget — decisive for production agents.
Long-Horizon Tasks	Sustained multi-step planning and memory over long-running tasks.

🔑 Mind Map (Framework Overview)

📥 Download High-Resolution Mind Map (mindmap.png)

🔍 Click here to unfold the full Mind Map
(点击展开完整思维导图)

---

📑 Presentation Slides

A comprehensive guide to evaluating AI agents, focusing on the engineering framework for testing — including the "Clean Room" methodology, reliability metrics (Pass@k), and the "Harness" architecture. It treats evaluation as a core development practice.

💡 Tip: Press Ctrl + Click (or Command + Click) to open in a new tab.
📥 View Slides (PDF)
📥 Download PDF (Direct Link)

---

🧰 Key Tools, Frameworks & Strategies

1. The Tooling Stack (Ecosystem)

Implementing a robust evaluation pipeline requires specific infrastructure. The following tools are referenced and utilized in this framework:

Tool	Category	Key Features
LangSmith	Tracing & Debugging	Full trajectory tracing, runnableConfig tagging for A/B testing, and dataset management.
LangFuse	Observability	Open-source alternative for observability, prompt management, and lightweight evaluation.
DeepEval	Unit Testing	"Pytest for LLMs". Specific metrics for RAG (Hallucination, Answer Relevancy) and Agents.
OpenEvals	Graders	A library of pre-built "LLM-as-a-judge" prompts (Conciseness, Correctness, Coherence) compatible with LangSmith.
AgentOps	Agent DevOps & Monitoring	Session replays, agent benchmarking, and cost & reliability tracking for autonomous agents.
Promptfoo	Red Teaming & Regression	Declarative eval configs, LLM regression testing, and adversarial red-teaming.
Braintrust	Evaluation Platform	Dataset management, LLM-as-a-judge scoring, online evals, and A/B testing.
Arize Phoenix	Observability & Eval	Open-source LLM tracing, embedding analysis, and RAG/agent evaluation.
W&B Weave	Tracing & Eval	Lightweight LLM instrumentation, eval harnesses, and dataset versioning.
OpenAI Evals	Evaluation Framework	OpenAI's open-source framework for model-graded evals — YAML/JSON config-driven, custom eval classes, and dataset registries.

2. Architecture: Hybrid Agent (Fast vs. Slow)

To balance cost and performance, we implement a Hybrid Agent Architecture:

• Reactive Layer (System 1): Handles simple, direct queries (e.g., "What is the stock price?") with low latency.
• Deliberative Layer (System 2): Activated for complex planning or multi-step reasoning tasks.
• Coordination Layer: A router that classifies intent and dispatches tasks.

Each layer is evaluated with different metrics:

• Reactive Layer: latency and single-step accuracy.
• Deliberative Layer: task completion rate, multi-step planning correctness, and trajectory quality.
• Coordination Layer: intent classification accuracy — misrouting is a common source of downstream failures.

3. Evaluation Strategy: The "Clean Room"

To prevent "cheating" through shared state, every evaluation trial runs in a fresh container/sandbox.

• Isolation: Fresh container for every trial, plus state reset (environment, conversation history) and snapshot rollback to guarantee a clean slate.
• Mocking & Replay: Simulate external APIs — or record-and-replay real responses — to control latency and produce deterministic, reproducible outputs.
• Determinism: Fix seeds and use temperature=0 so runs are repeatable and differences are attributable to code, not randomness.
• Cleanup & Anti-Leakage: Aggressive state teardown (no shared history) and guard against goal leakage that could let the agent take shortcuts via shared state.

⬆️ Back to Top : Table of Contents