6 Weeks
Every week you ship something real. By week 6 you have a complete, production grade AI system with real world reach, documented, and yours to own.
Reality Check & First Ship
Theme: Feel it before you engineer it.
Most bootcamps start with theory. We start with a working system in production. By Sunday you've shipped a real CrewAI powered tool. That experience + the excitement
LLMs as stateless text transformers ; tokens, context windows, the agent loop.
Hallucination, runaway spend, latency, no fallback, prompt drift. Name them before you code.
tokens × price per token × call frequency = your bill. Calculate it before writing a line.
Multi agent system in ~50 lines. Use it deliberately ; feel the ceiling, then understand why.
- AI Reality Map ; system diagram with named failure modes
- Market Scout ; live public URL (CrewAI + FastAPI)
- Cost report ; actual spend + projected monthly
Market Alignment
Theme: Same behaviour. No framework. Full control.
Week 1 gave you a working system fast. Week 2 answers: what was CrewAI actually doing? Rebuild Market Scout using raw OpenAI API calls and explicit orchestration ; same inputs, same outputs, full visibility. Then pivot: find a real market need, talk to one user, and ship a demo they can touch.
Write the loop that calls the model, parses the response, decides the next step. No magic.
Track cumulative spend per request. Hard stop at a ceiling. Log every token.
Use response_format with JSON schema. Enforce structure at the API level.
Skill inventory → problem intersection → user interview → targeted demo. Ship by Sunday.
- Market Scout v2 ; raw API rebuild, live public URL
- Cost comparison: Week 1 vs Week 2, annotated
- One page market brief + user interview notes
Tooling & RAG
Theme: Your demo has no memory. Fix that.
The Week 2 demo impresses users ; until they ask a follow up. The model has no context from the previous exchange, no access to your data, no way to ground answers in facts. You'll build a retrieval augmented generation pipeline from scratch, evaluate the ecosystem, and make explicit buy vs. build decisions.
Turn text into vectors. Store in pgvector. Query by semantic similarity, not keyword.
Embed query → retrieve top k chunks → inject context → generate. Own every step.
Fixed size vs. sentence aware vs. semantic. How you split determines retrieval quality.
Pinecone, Weaviate, pgvector, LlamaIndex ; evaluated against your actual requirements.
- RAG enabled system ; live, grounded, with citation logging
- Chunking experiment report ; method comparison + scores
- Buy vs. build matrix ; every layer, with rationale
AI Pipelines & Agents
Theme: Move AI off the request path.
Event driven architecture, graph based agent orchestration, MCP tools, and vector search wired together into a full end to end pipeline. The model stops being a synchronous API call and becomes a durable background process that reasons, acts, and stores results.
Trigger AI jobs from events. Handle them durably. Keep your API fast.
ReAct loops, graph based orchestration, worker + tool nodes wired into reliable pipelines.
Model Context Protocol: build a tool, register it, let your agent discover and call it.
Embeddings + pgvector integrated into the pipeline. Agent retrieves relevant context on demand.
- End to end AI pipeline: event → agent → tool calls → stored result
- One working MCP tool registered to your agent
- Technical walkthrough of your system architecture
Reliability & Deployment
Theme: Works on my machine → runs in production.
Durable execution, step by step tracing, container setup, and launch prep. Your Week 4 pipeline is impressive ; now make it survive failures, give you full visibility into what happened, and run somewhere that isn't your laptop.
Retries, resumable steps, idempotent design. Your pipeline survives failures without duplicating work.
Trace agent runs step by step. See exactly what happened, what state each node produced, where it failed.
Container setup, environment config, services talking to each other in a real environment.
Repo structure, README, setup instructions. If it takes >10 minutes to clone and run, it's not ready.
- Hardened pipeline ; retry safe, failure tolerant
- Execution traces ; visible, step by step agent run logs
- Live deployed project + public repo, clone and run ready
Build Something That Matters ↗
Theme: You choose the problem. You own the outcome.
No spec. No prescribed stack. Five weeks of foundations, retrieval, pipelines, and hardening — now apply them to something you actually care about. Pick a real problem, build a working solution using whatever combination of tools makes sense, and present it to the cohort.
Define a problem worth solving. One sentence. Who has it, what it costs them, why AI is the right lever.
Ship a working system. It should be live, usable, and reflect at least three weeks of curriculum. Your decisions, your tradeoffs, your code.
Present to the cohort. Walk through the problem, the system, one live demo, and what you'd do differently. 10 minutes. No slides required.