tensorfolk

Learn AI engineering

54 topics in 12 tracks, from how models work to running them in production. Each topic is short: the key ideas, the common mistakes, and where to go deeper. Last reviewed 17 Oct 2026.

LLM Foundations

How large language models work, from tokens to training stages, and how to pick one.

  1. Transformer architectureStableAlmost every modern LLM is a decoder-only transformer: a stack of self-attention and feed-forward blocks that predicts the next token from all previous tokens.
  2. Tokenization and context windowsStableModels read and bill in tokens, not words. The context window is the hard limit on input plus output tokens per request.
  3. Decoding and samplingStableThe model outputs a probability distribution for the next token; the decoding strategy decides which token is picked.
  4. EmbeddingsEvolvingEmbedding models map text (or images) to dense vectors so that semantic similarity becomes geometric distance. They power search, clustering, dedup and RAG.
  5. How models are trainedEvolvingPretraining builds raw capability; post-training turns it into a helpful, tool-using, reasoning assistant.
  6. Model landscape and selectionChanges fastModel rankings change monthly. Select models with your own evals against quality, latency, cost, context, tool use and deployment constraints.

Prompt & Context Engineering

Getting reliable behaviour from a model through instructions, examples, structure and well-managed context.

  1. Core prompting patternsStableClear, specific instructions with context, examples and an explicit output format beat clever tricks.
  2. Structured outputsEvolvingWhen code consumes model output, constrain it to a schema and validate it.
  3. Context engineeringEvolvingDeciding exactly which tokens the model sees at each step: instructions, retrieved data, tool results, memory and history. Context is a finite budget.
  4. Prompt cachingEvolvingProviders can cache a repeated prompt prefix so later calls skip reprocessing it, cutting input cost and latency substantially.
  5. Reasoning and effort controlsChanges fastReasoning models expose knobs (thinking budgets, effort levels) that trade latency and cost for accuracy.

Retrieval & RAG

Grounding model answers in your own data with search, chunking, reranking and evaluation.

  1. RAG pipeline end to endStableRetrieval-augmented generation fetches relevant passages at query time and puts them in the prompt so the model answers from your data and can cite it.
  2. Chunking strategiesEvolvingHow you split documents decides what can be retrieved. Chunks must be small enough to be precise and large enough to be self-contained.
  3. Vector stores and indexesEvolvingVector databases store embeddings and run approximate nearest-neighbour (ANN) search with metadata filters.
  4. Hybrid search and rerankingStableCombine keyword search (BM25) with vector search, then rerank the merged candidates with a cross-encoder. This is the highest-leverage RAG upgrade.
  5. Advanced and agentic RAGChanges fastTechniques beyond single-shot retrieval: query transformation, graph-based retrieval and agents that search iteratively.
  6. Evaluating RAGEvolvingMeasure retrieval and generation separately so you know which part to fix.

Agents & Tool Use

LLMs that call tools in a loop to complete multi-step tasks, and the patterns that keep them reliable.

  1. The agent loopStableAn agent is a model in a loop: it reads context, decides to call a tool or answer, receives tool results, and repeats until done.
  2. Designing tools for agentsEvolvingTool quality drives agent quality. Design tools for a model reader: clear names, rich descriptions, small focused inputs and concise outputs.
  3. Workflow patternsStableMany production systems are workflows with fixed code paths that call LLMs, not free-roaming agents. Pick the simplest pattern that works.
  4. Agent memoryChanges fastMemory lets agents carry information across turns and sessions: conversation history, user facts and learned procedures.
  5. Multi-agent systemsChanges fastSeveral agents with separate contexts and roles cooperate, usually a lead agent delegating to sub-agents.
  6. Agent frameworksChanges fastFrameworks give you the loop, tool schemas, state and tracing. Choose by how much control you need.

Protocols & Integration

Standard ways to connect models to tools, data and other agents.

  1. Model Context Protocol (MCP)EvolvingMCP is an open JSON-RPC protocol that lets any AI host connect to tool and data servers through one interface. Now governed by the Agentic AI Foundation.
  2. MCP 2026-07-28 spec: stateless coreChanges fastThe 2026-07-28 revision removes the initialize handshake and sessions so any server instance can answer any request, and adds Tasks and MCP Apps extensions.
  3. Agent-to-agent (A2A)Changes fastA2A is an open protocol for agents from different vendors to discover each other and delegate tasks. MCP connects agents to tools; A2A connects agents to agents.
  4. Auth for agents and toolsEvolvingAgents act on behalf of users, so every tool call needs a clear identity, scoped permissions and an audit trail.

Evaluation

Knowing whether your AI system works, and catching regressions before users do.

  1. Eval strategyStableEvals are the unit tests of AI systems. Start from real failure modes, write targeted checks and run them on every prompt, model or tool change.
  2. LLM-as-judgeEvolvingA model grades outputs against a rubric. Cheap and scalable, but it must be calibrated against human labels.
  3. Evaluating agentsChanges fastAgents need evals of the final outcome and of the path: tool choice, arguments, number of steps and cost.
  4. Public benchmarksChanges fastBenchmarks signal general capability but saturate fast. Use them to shortlist models, not to choose one.
  5. Eval and tracing toolsChanges fastTools to store datasets, run experiments, trace production calls and review outputs.

Fine-tuning & Adaptation

When and how to change model weights instead of prompts.

  1. Prompt vs RAG vs fine-tuneStablePrompting changes instructions, RAG changes knowledge at query time, fine-tuning changes behaviour. Try them in that order.
  2. LoRA, QLoRA and PEFTStableParameter-efficient fine-tuning trains small adapter matrices instead of all weights, cutting memory and cost dramatically.
  3. Preference tuning and RLEvolvingAfter SFT, preference methods teach a model which of two outputs is better; RL with verifiable rewards teaches it to solve checkable tasks.
  4. Distillation and synthetic dataEvolvingUse a strong model to generate or label data, then train a smaller, cheaper model to match it on your task.

Inference & Serving

Running models fast and cheaply, whether through an API or on your own GPUs.

  1. Inference fundamentalsStableInference has two phases: prefill processes the prompt in parallel, decode generates tokens one at a time. Most performance work targets one or the other.
  2. Serving enginesChanges fastOpen-source engines that serve open-weight models with an OpenAI-compatible API.
  3. QuantisationEvolvingStoring weights (and sometimes activations and KV cache) in fewer bits to cut memory and speed up inference.
  4. Cost and latency optimisationEvolvingMost savings come from sending fewer tokens to smaller models, and from caching.

Production & LLMOps

Observability, safety, security and reliability for AI systems in production.

  1. Observability and tracingEvolvingTrace every model and tool call with inputs, outputs, tokens, latency and cost so you can debug, evaluate and bill.
  2. Security and prompt injectionChanges fastAny text the model reads (web pages, emails, documents, tool output) can contain instructions. Design so that a manipulated model cannot cause serious harm.
  3. GuardrailsEvolvingChecks around the model that block, rewrite or escalate inputs and outputs that break policy.
  4. Reliability patternsStableLLM APIs fail, throttle and slow down. Build the same resilience you would for any external dependency.
  5. Governance and regulationChanges fastData handling, transparency and risk rules that shape enterprise AI deployments.

Cloud AI Platforms

Managed services for models, agents and retrieval on AWS, Azure and Google Cloud.

  1. Amazon BedrockChanges fastAWS's managed access to foundation models (Claude, Llama, Mistral, Nova and more) with retrieval, guardrails and fine-tuning around them.
  2. Amazon Bedrock AgentCoreChanges fastAWS building blocks to run agents built with any framework: runtime, memory, gateway, identity, code interpreter, browser and observability.
  3. Azure and Google CloudChanges fastEquivalent stacks on the other two hyperscalers.

Multimodal & Voice

Working with images, documents, audio and screens.

  1. Vision and document understandingEvolvingVision-capable models read images, charts, scans and PDFs directly, often replacing classic OCR pipelines.
  2. Voice agentsChanges fastVoice agents either chain speech-to-text, LLM and text-to-speech, or use a native speech-to-speech realtime model.
  3. Computer and browser useChanges fastAgents that operate GUIs through screenshots, clicks and typing, for systems without APIs.

AI System Design

Putting the pieces together into architectures you can defend in a design review or interview.

  1. Reference architecture: enterprise copilotEvolvingA typical copilot combines chat UI, an agent service, tools over company systems, retrieval, memory and an eval and observability layer.
  2. Multi-tenancyStableIsolating data, cost and limits between customers in a shared AI platform.
  3. Approaching AI design questionsStableA repeatable structure for AI system design reviews and lead-level interviews.