ShipAI
← Back to Main
Learn AIFull curriculum outline

Concepts

Foundations → deeper theory

Core Concepts
  • What is AI (and what it is not)
  • Classical ML literacy
  • Tokenization
  • Self-attention
  • Embeddings and similarity
  • Prompt engineering fundamentals
  • Structured outputs
  • RAG building blocks
  • Agents and the ReAct loop
  • Evals fundamentals
  • Serving and streaming
  • Networking for AI apps
  • LLM project lifecycle
  • Multimodal basics
  • Privacy and data for AI apps
  • Open-weight models vs closed APIs
Advanced Concepts
  • Scaling laws and compute
  • Fine-tuning with LoRA and QLoRA
  • Alignment basics — RLHF and DPO
  • Mixture of Experts (MoE)
  • Context engineering
  • Reasoning and test-time compute
  • Multi-agent orchestration
  • Guardrails and safety systems

Tools

Everyday stacks → production tooling

Key Tech
  • LangGraph and LangChain patterns
  • Hugging Face
  • Ollama
  • Model Context Protocol (MCP)
  • LlamaIndex
  • OpenAI and Anthropic APIs
  • Chroma
  • Weights & Biases
Advanced Key Tech
  • Ray
  • Triton Inference Server
  • Kafka for evented AI
  • OpenTelemetry for LLMs
  • TensorRT-LLM and SGLang
  • MLflow for LLMOps

Data & inference

Retrieval data plane + serving

Data & Databases for AI
  • Vector databases — what, why, and how
  • Choosing vector stores — Pinecone, Weaviate, Chroma, pgvector
  • Postgres and pgvector
  • Redis for AI caching
  • Chunking and metadata
  • Hybrid search and rerankers
Inference
  • vLLM
  • KV-cache, prefill, and decode
  • Continuous batching
  • Quantization for inference
  • Speculative decoding
  • Cost and latency routing

Guided path

Projects · foundations → build → agents → production

Foundations

Onboarding
  • How this course works
  • Python, tooling, and compute options
  • API keys, budgets, and responsible use
  • Rubric: tests, eval JSON, and README
What happened in AI (till now)
  • Symbolic AI → why rules hit a wall
  • Classical ML era
  • Deep learning boom
  • Attention → Transformer → GPT lineage
  • Generative AI landscape
  • Open-weight vs closed API
  • Jobs: AI eng, ML eng, research, FDE, PM
ML/DL literacy
  • Data splits and leakage
  • Loss, gradients, and overfitting
  • Embeddings before LLMs
  • Autograd intuition
  • PyTorch training loop
Build an LLM from scratch
  • Tokenization and BPE
  • Next-token prediction
  • MLP language model
  • Self-attention
  • Tiny GPT / mini-LLM end-to-end
  • Sampling (temperature, top-k/p)
  • Pretrain vs SFT vs preference (map only)
  • Why your tiny LLM ≠ ChatGPT

Generative AI practice

Talk to models in the real world
  • Chat APIs and message roles
  • Structured output
  • Prompt versioning
  • Multimodal in → text
  • Code assist loops
RAG
  • Embeddings and similarity
  • Chunking
  • Vector DB
  • Hybrid search
  • Rerank / rewrite
  • Citations and failures
Build & serve your SLM
  • When SLM/FT beats RAG or prompting
  • Dataset curation and licenses
  • LoRA / QLoRA fine-tune
  • Eval vs base / teacher
  • Serve your SLM
  • Quantization tradeoffs

Agents & agentic systems

Build real AI agents
  • LLM vs agent; loops; stop conditions
  • Tools (schema, validation, side effects)
  • Planning vs reactive
  • Memory
  • Human-in-the-loop
  • Trace and debug trajectories
  • Framework packaging
Skills, MCP, context engineering
  • Skill vs tool vs prompt
  • Author skills
  • Dynamic skill loading
  • MCP servers and clients
  • Context budgets
  • Coding-agent skill patterns
Agentic workflows & multi-agent
  • Workflow patterns
  • Shared state and handoffs
  • Agentic RAG
  • Long-running / checkpoint / resume
  • Failure isolation
  • Harness thinking

Production + how companies actually do it

Evals, guardrails, safety
  • Offline vs online evals
  • LLM-as-judge
  • Component vs E2E
  • Injection / jailbreaks
  • Guardrails
  • Canaries
Deploy, cost, latency, observability
  • API + streaming
  • Cost accounting
  • Caching and latency
  • Tracing tool/LLM spans
  • Retries, rate limits, idempotency
  • Runbook
How real companies use AI
  • Netflix — in-house LLM serving
  • Uber — agent platform, MCP, gateway
  • Meta — LLM-scale training literacy
  • Stripe — coding agents / harness
  • Shopify — production merchant agent
  • Airbnb — verify-heavy workflows
  • NVIDIA — agent toolkits and eval/obs
  • Google — agentic training / RL direction
AI system design
  • Build vs buy
  • UX of uncertainty, citations, feedback
  • Data flywheels
  • Privacy / retention sketch
  • Capstone proposal

Frontier + capstone

What’s coming + staying current
  • Reasoning models / test-time compute
  • Multimodal agents
  • On-device / small models
  • Synthetic data loops
  • Safety, regulation, dual-use
  • Personal radar
Capstone
  • Capstone requirements and pick-two
  • Architecture and eval plan
  • Build and harden
  • Write-up and demo
Sign up / in
Browse tracksInference

Learn AI · Browse tracks

Inference

Serving stack — KV-cache, continuous batching, quantization, speculative decoding, vLLM, cost and latency routing.

How this relates Browse here for Hello Interview–style articles. The guided path below in the sidebar remains the project milestone sequence.

Articles in this track

  • 1. 90 minvLLMHigh-throughput LLM serving — PagedAttention, continuous batching, scheduling knobs, capacity planning, and when vLLM beats naive Hugging Face generate.Open
  • 2. 75 minKV-cache, prefill, and decodeWhy TTFT and tokens/sec are different jobs — KV growth math, prefix caching, chunked prefill, and scheduling implications for multi-tenant serving.Open
  • 3. 70 minContinuous batchingStatic batches waste GPU on short requests; continuous batching admits new sequences as others finish — scheduler knobs, fairness, and prefill/decode mixing.Open
  • 4. 75 minQuantization for inferenceAWQ, GPTQ, GGUF, FP8 — shrink weights/KV to raise concurrency; calibrate on prod-like data and measure task quality, not only VRAM.Open
  • 5. 70 minSpeculative decodingDraft tokens with a small model, verify with the large one — speedups when acceptance rates stay high; when to enable per route.Open
  • 6. 75 minCost and latency routingRoute easy traffic to cheap/fast models and hard traffic to frontier — with budgets, caches, SLOs, escalation, and FinOps logging.Open
NextvLLM

ShipAI curriculum

A structured path from foundations to production agents — lessons, projects, and milestones in one course. Free forever.

View course
Reading progress

On this page

  • Overview
  • Articles
  • vLLM
  • KV-cache, prefill, and decode
  • Continuous batching
  • Quantization for inference
  • Speculative decoding
  • Cost and latency routing