The advanced track

What you'll build

• 5-stage NetOps text pipeline using only LCEL — beats imperative loops 5-10x via auto-batching
• 10 NetOps tools converted to LangChain + MCP bridge to your existing servers
• Hybrid retriever (BM25 + dense + reranker) — beats vector RAG by 10pp on MRR
• Hand-built ReAct agent as a StateGraph — checkpointer, HITL, streaming all native
• LangSmith A/B test: two prompt variants, 3 evaluators, statistical-significance comparison
• LangServe deployment: /invoke, /batch, /stream, /playground, typed client SDK — all free
• LangMem-powered long-term memory: per-operator preferences across sessions
• C1 crew ported to a LangGraph Supervisor — matches CrewAI eval within ±5pp

What you walk away with

A migration playbook for your team's wiki: "when LangChain wins, when CrewAI wins, when both." Plus working code for all 9 lessons + the LC1 capstone — the framework-judgment artifact a hiring manager actually wants to see.

What you'll build

• Lifecycle-instrumented crew with step_callback + task_callback — flame-graph the agent execution
• Custom BaseTool subclass with caching, retry, Pydantic args_schema
• Event-driven Flow with @start/@listen/@router — replaces sequential workflows that branch
• 4-tier memory: short-term, long-term, entity, user — Chroma-backed
• AgentOps wiring + custom cost-tracking callback — identify the most expensive step in any run
• VCR-based regression test suite — deterministic replay, CI gate on 10% quality drop
• FastAPI deployment + multi-stage Dockerfile + Kubernetes manifest
• Triage→Remediation multi-crew architecture with Flow-orchestrated handoff

What you walk away with

Deep CrewAI mastery beyond the M5 basics. The CR1 capstone: a production-shaped multi-crew system you can demo, plus the memo on "when CrewAI Enterprise SaaS pays off."

What you'll build

• FastMCP server exposing 4 tools + 2 resources — composite RCA + 3 atomic primitives
• CrewAI agent crew that consumes the MCP server via langchain-mcp-adapters
• 20-scenario eval comparing direct Python imports vs MCP vs MCP+cache
• 1-page memo: "When does our team adopt MCP for retriever access?"

Why it matters

"My retriever is a Python import; no other agent can call it." The C1 deliverable turns your retriever into a tool any agent on any framework can consume. The eval table — same accuracy, ~10x latency cost, swap-frameworks-for-free architectural win — is the artifact a hiring manager understands.

What you'll build

• FastAPI gateway with 5-tenant policy matrix — per-tenant quotas, cache namespaces, tier policies
• Tiered routing: heuristic classifier + tenant policy + automatic failover via circuit breakers
• 100-operator Locust load test with realistic 60/30/10 Tier-1/2/3 traffic mix
• Routing-strategy comparison: all_haiku / heuristic / llm_router / all_opus across 20 scenarios
• 1-page memo: "What's our tier policy for each business unit?"

Why it matters

"Tenant A's cache hit just answered tenant B's question." Multi-tenancy is the production failure mode no one talks about until it bites. This capstone forces you to design isolation boundaries deliberately — auth, quota, cache, routing — and measure the cost-quality-latency triangle for each tenant.

What you'll build

• QLoRA fine-tune of Llama-3.1-8B on 5,000 incident summaries — served via Ollama as a local_ft gateway tier
• Opik traces spanning gateway → CrewAI agent → MCP tool call → backend LLM (one trace ID end-to-end)
• CI eval gate that runs on every PR — blocks merge on quality regression past threshold
• Streamlit cost/quality/latency dashboard for the platform owner
• 8-failure-mode on-call runbook (tier outage, cost spike, drift, eval failure, model gibberish, MCP crash, recurring symptom, spend-cap)
• 1-page memo: when does this stack make sense, when is it over-engineering, biggest risk

Why it matters

"I have a notebook and no story about cost, drift, or rollback." This is the curriculum's portfolio-grade artifact — the thing graduates demo to hiring managers and architectural review committees. Code shows you can build; the memo shows you can decide; the runbook shows you can operate.