Summary of Core AI Concepts that I will be Covering
In terms of the plan to cover the breadth of AI key topics to provide some level of structure to the posts, please see the following list. The aim is to develop a high level understanding of the key concepts and then build the depth. This will help ground the direction and understanding of fundamental concepts, so that as detail is added, there is still a strong foundation in place and understanding of where it sits in the bigger picture.
In terms of the core topics that will be covered, my mental model uses the Large Language Model as the brain and agentic AI as the remainder of the body, which ultimately enables the action.
What does this mean? This means that I will structure the posts as LLM, which will focus on ensuring that the brain is best used, understood and does not degrade over time. I have captured a set of potential sub-topics that may come up during the blog if relevant. From the agentic perspective, the aim is to ensure that the latest considerations are taken onboard as this is the technology that I will be deploying.
SO LLM topics are
- LLM Foundations
- LLM APIs
- Open Source LLMs
- Prompt Engineering
- Structured Output
- Vector Stores
- Embedding Models
- Vector Databases
- Chunking Strategies
- Semantic Search
- RAG
- Ingesting docs
- Retrieval methods
- Context handling
- Prompts Templates
- Orchestration Framework
- Advanced RAG
- Query transformation
- Reranking & Filtration
- LLM as a judge
- HyDE
- Corrective RRAG
- RAGAS Evaluation
- Agentic RAG
- Graph-RAG
- Self-RAG
- Fine Tuning
- Data Preparation
- PEFT Methods
- Training Config
- Alignment
- Training Tools
- Inference Optimisation
- Quantisation
- Serving Engine
- Optimisation Techniques
- Deployment
- MML/LLMOps
- Infrastructure
- Local inference cloud platforms
- Observability
- Tracing and logging
- Monitoring metrics
- Evals
- Agents high level (intro)
- Why
- Landscape
- Production & Security
- Guardrails
- Cost optimisation
- Reliability
Agentic AI
- Agent Foundations
- What is an agent vs. a chain vs. a workflow
- The ReAct pattern (Reason + Act)
- Tool use and function calling
- Memory types (short-term, long-term, episodic)
- Tool & Environment Integration
- API tool design and schemas
- Code execution sandboxes
- Browser/web interaction
- File system and database access
- Planning & Reasoning
- Chain-of-thought and tree-of-thought
- Task decomposition strategies
- Self-reflection and critique loops
- Goal tracking and replanning
- Orchestration Frameworks
- LangGraph
- CrewAI
- AutoGen
- OpenAI Agents SDK
- Anthropic's tool use patterns
- Multi-Agent Systems
- Agent roles and delegation
- Communication protocols between agents
- Supervisor vs. peer architectures
- Shared state and handoffs
- Memory & State Management
- Conversation memory strategies
- Vector store integration for long-term recall
- Context window management
- Checkpointing and resumability
- Human-in-the-Loop
- Approval gates and breakpoints
- Escalation patterns
- Confidence thresholds for autonomous action
- User feedback loops
- Retrieval-Augmented Agents
- Agents that decide when to retrieve
- Dynamic tool selection
- Combining RAG with planning
- Agentic RAG vs. static RAG
- Evaluation & Testing
- Trajectory evaluation (did the agent take good steps?)
- End-to-end task success metrics
- Benchmarks (SWE-bench, GAIA, WebArena)
- Regression testing for agent behaviour
- Safety & Guardrails
- Action sandboxing and permissions
- Prompt injection defence
- Output validation before execution
- Rate limiting and cost controls
- Deployment & Observability
- Tracing agent execution (LangSmith, Arize, etc.)
- Latency and cost monitoring per step
- Error recovery and fallback strategies
- Logging decisions and tool calls
- Production Patterns
- Async and parallel agent execution
- Caching strategies for repeated tasks
- Graceful degradation when tools fail
- Versioning agent configurations