8 weeks. 3 phases.
Everything you need to build.

Transformer architecture conceptually — attention, tokens, context windows. Why models predict next tokens. Hallucination: what it is, why it happens, how to design around it. Live demo: instructor breaks a prompt on purpose and the class diagnoses why. Students write their first professional system prompt for their domain.

Chain-of-thought prompting. Few-shot examples. Structured JSON output. Output format control. Students rebuild their Monday system prompt with all three techniques, then test 10 inputs and score each on accuracy, format, and usefulness.

Why evals are the most underrated skill in AI development. The difference between a demo that works and a product that works. Students build a 10-case eval set with hand-written ideal outputs, score their current prompt, and name their top failure mode.

The no-code AI landscape. When to use which tool. The mental model: you're the product manager, the AI is the engineer. Live demo: instructor builds a complete, deployed AI tool in Lovable in 30 minutes. Students build and deploy their own using their product brief.

What an API is and why it matters. Making a direct Claude API call from Replit. API cost awareness — calculating cost per call and extrapolating to 100 users. Connecting Lovable front ends to real API backends. Error handling so users see a useful message when something breaks.

Students present their first deployed domain-specific AI tool. 3 minutes each: what it does, a live demo, what breaks and why, and what you'd fix first. Written feedback from instructor and cohort. This is the baseline everything else builds on.

What RAG is and why it matters. The pipeline: query → embed → search → retrieve → inject → generate. Vector databases and semantic search. Students set up a Pinecone index, embed 10 domain documents, and run their first semantic queries.

Chunking strategies and metadata. Ingesting 50+ real domain documents with a repeatable script. Hybrid search (semantic + keyword). Metadata filtering. Reranking. Students identify their top retrieval failure mode and implement the fix.

Measuring retrieval precision vs. generation groundedness. Building a RAG eval set. Students build the full Lovable front end: chat interface, citations, out-of-scope handling, streaming responses. URL exchanged by end of session for peer feedback.

Streaming responses so users see output in real time. Fallback messages for pipeline failures. Query logging to Google Sheets or Airtable. Semantic caching for common queries. Students implement all four and run a production readiness check.

The think-aloud protocol. In-class user testing pairs. Students identify their top 3 user friction points and prioritize the single highest-impact fix before the deliverable. External user sessions scheduled for this week.

Students present their deployed RAG product. 4 minutes: the data it knows, a live demo, user testing findings, and top failure mode. Instructor and LATAM team provide full written assessment by Saturday morning.

The agentic loop: perceive, think, act, observe, repeat. Tool use in Claude. Single-tool agents. Multi-tool agents with domain-specific tools. Reliability and defensive design — human-in-the-loop checkpoints, confirmation steps for write actions, stress testing completion rates.

Mapping a real business workflow. Event-driven automation: form submission → AI reasoning → route to owner → log to Airtable. Multi-agent orchestrator/subagent patterns for complex tasks. Students build the first version of their Week 5 automated workflow.

Students present their deployed AI agent. 5 minutes: the workflow it automates, a live end-to-end demo, stress test results (completion rate + top failure mode), and what they'd build next. LATAM team full assessment by Saturday evening.

Connecting AI to a real Postgres database. When AI should read vs. write. Row-level security so users only see their own data. Students wire a Supabase database to their capstone — reads inform responses, writes require user confirmation.

Real-time subscriptions. Scheduled ingestion jobs. Students present their capstone architecture on the whiteboard — cohort identifies what will break under load and what has no fallback. Live data pipeline tested end to end.

The production checklist: error handling, fallbacks, logging, security, cost instrumentation, input sanitization, rate limiting. Live audit of a volunteer capstone. Students complete their own audit, document every gap with priority, and fix the top three issues.

Onboarding, empty states, user-friendly error messages. The stranger test: a stranger uses your tool without explanation — you watch and take notes without helping. Students fix their top 3 productization gaps and explicitly scope v1.

Positioning your AI product. Pricing models: per-seat, usage-based, outcome-based. Finding your first 10 customers. The one-page GTM plan. 90-second pitches to the cohort — no deck, just talking. Students write their full GTM plan and get hard feedback from peers.

API cost modeling at scale. Model selection trade-offs. Prompt caching and compression. Students build a cost model showing margin at 10, 100, and 1,000 users. Demo Day format briefed. Full 7-minute script written and peer-reviewed.

No lectures. Full build time. Instructor and LATAM team on floor for 1:1 support. 1:1 office hours with instructor — 5 minutes each, one thing to complete before Demo Day. Every capstone production-ready by end of day.

Room set up exactly like Demo Day. Every student delivers the full 7-minute presentation. Timed strictly. Written feedback immediately after each. Second run for students with significant changes. Final logistics, Q&A prep, contingency plans.

7 minutes per student. Live audience of Austin operators, investors, and hiring partners. Every student presents a production-ready AI product with a live demo. Q&A from domain experts and investors. Documented outcomes shared publicly.