AI Advanced

Build multi-agent systems using Google ADK, LangGraph, and other orchestration frameworks.

• Google Agent Development Kit (ADK): agents, tools, sessions, runners
• ADK agent types: LlmAgent, SequentialAgent, ParallelAgent, LoopAgent
• ADK tools: FunctionTool, AgentTool, built-in tools (code execution, search)
• ADK memory: in-session, persistent (database-backed), session service
• LangGraph: state graphs, nodes, edges, conditional routing, cycles
• LangGraph persistence: checkpointers (SQLite, Postgres), human-in-the-loop
• Multi-agent patterns: orchestrator-worker, supervisor, parallel fan-out
• Agent communication: handoffs, shared state, message passing
• CrewAI: crews, agents, tasks, tools, hierarchical process
• Evaluating agent systems: trajectory evaluation, task success rate, cost per task

Free Agent Frameworks & ADK practice questions →Agent Frameworks & ADK study guide →

Build production applications with the Anthropic SDK, tool use, extended thinking, and MCP.

• Anthropic Python/TypeScript SDK: client setup, messages.create, streaming
• Tool use (function calling): tool definitions, input_schema, tool_result blocks
• Parallel tool use: multiple tool calls in one turn, result aggregation
• Extended thinking: thinking budget, streaming thinking blocks, when to enable
• Computer use: screenshot/click/type tools, security considerations
• Batch API: message batches, async polling, large-scale inference
• Prompt caching: cache_control breakpoints, cache hit/miss, TTL
• Model Context Protocol (MCP): server authoring, Claude Desktop config
• Claude Code SDK: spawning subagents, reading files, tool permissions
• Anthropic safety features: Constitutional AI, harmlessness training background

Free Claude SDK & Anthropic APIs practice questions →Claude SDK & Anthropic APIs study guide →

Instrument AI applications for tracing, evaluation, and continuous quality improvement.

• Why observability matters for AI: non-determinism, prompt drift, silent failures
• OpenTelemetry for AI: tracing LLM calls, spans, attributes (model, tokens, latency)
• Langfuse: traces, spans, generations, scores, dashboards — open-source LLM observability
• Braintrust: experiments, datasets, scoring functions, CI integration
• Arize AI and Phoenix: model monitoring, embedding drift, retrieval quality
• Structured logging for AI: what to log (prompt, response, tokens, latency, cost)
• Regression testing: golden datasets, CI eval runs, alert on score drop
• A/B testing prompts: experiment design, statistical significance
• Cost attribution: per-request cost tracking, budget alerts
• Human feedback loops: thumbs up/down, annotation tools, RLHF data collection

Free AI Observability & Evals practice questions →AI Observability & Evals study guide →

Design resilient, cost-efficient AI systems for production — streaming, caching, self-hosted models, and scale.

• Streaming UX: SSE in Next.js API routes, AI SDK useChat, token-by-token rendering
• Semantic caching: caching LLM responses by embedding similarity (GPTCache)
• Prompt caching at provider level: Anthropic, OpenAI, Gemini — rules and savings
• Router patterns: model routing by cost/capability, LiteLLM, OpenRouter
• Fallback and retry: provider failover, exponential backoff, circuit breaker
• Self-hosted inference: vLLM serving, OpenAI-compatible API, GPU sizing
• Ollama in production: reverse proxy, authentication, multi-model serving
• Context window management: conversation pruning, summarization, sliding window
• Rate limit management: token bucket, queue-based throttling, user quotas
• Cost modeling: estimating monthly LLM spend, optimization levers

Free Production AI Architecture practice questions →Production AI Architecture study guide →

Apply responsible AI principles, defend against AI-specific attacks, and meet compliance requirements.

• OWASP LLM Top 10: prompt injection, insecure output handling, training data poisoning
• Prompt injection attacks: direct vs indirect injection, jailbreaks, mitigations
• PII handling: detecting and redacting PII before sending to LLMs
• Output validation: structured output enforcement, hallucination detection
• Red-teaming AI systems: adversarial prompts, safety benchmarks, PAIR
• Responsible AI frameworks: Microsoft RAI, Google PAIR, EU AI Act overview
• Data privacy: no-training opt-outs, data residency, GDPR considerations
• Model cards and transparency: documenting intended use, limitations, biases
• AI incident response: handling model failures, rollback strategies
• AI governance tooling: Guardrails AI, NeMo Guardrails, Llama Guard

Free AI Safety & Governance practice questions →AI Safety & Governance study guide →