Your AI Startup, in 20 Days

The capstone is the AI product idea you've been carrying around for months — now built and shipped. A real product solving a real problem for real users who aren't doing you a favor. Public live URL with auth. Public eval scorecard. Public tech-blog write-up of one architecture decision. A 5-minute Loom that lands the impact in the first 30 seconds. A README that reads like senior-engineer documentation. This is the portfolio centerpiece that every other project artifact points back to.

The user persona is whoever your specific product idea targets. The constraint is that they must be real. Not classmates, not your mom. Real users you reached via cold outreach, Reddit, Hacker News, Twitter, a Discord community, or a tight-network ask. Onboarded with a working product. Provided feedback you actually acted on. Five of them, each using the product 3+ times. The north-star metric depends on your product — weekly active queries, task-completion rate, PR-author thumbs-up rate, or first-paid-customer for a micro-SaaS.

This is the project that converts a portfolio into job offers. Every Anthropic, Mistral, and Notion AI Engineer JD has the same line buried in it: "Production experience building LLM-powered applications." Many candidates have built side projects. Few have built side projects with real users, a public eval scorecard, and a published architecture write-up. Anthropic's own careers page says: "If you have done interesting independent research, written an insightful blog post, or made substantial contributions to open-source software, put that at the TOP of your resume."

You integrate every skill from Projects 1-5: RAG or advanced RAG (Project 1 or 4), an agent layer if relevant (Project 2), eval and observability instrumentation (Project 3), production hardening with guardrails, caching, and routing (Project 5). Cost-per-user-per-day must stay stable across a 7-day window. At least one feature ships because a user asked for it. At least one feature gets deprecated because users ignored it. This iteration loop — ship, measure, learn, ship — is the discipline that separates AI Engineers from AI experimenters.

What you ship: a live URL on your portfolio site. A 5-minute Loom that lands the impact in the first 30 seconds, then walks through the architecture. A one-page case study (problem, solution, metrics, lessons learned). A published tech-blog post about one architecture decision — with the data behind it. And a working product that continues to exist after the curriculum ends, that you can choose to keep operating, iterate on, or turn into a real business if the traction warrants it.

“Production experience building LLM-powered applications (prompting, context engineering, agents, evaluation frameworks, deployment)”

“What matters is what you ship, not the time you spend on it”

“If you have done interesting independent research, written an insightful blog post, or made substantial contributions to open-source software, put that at the TOP of your resume.”

“Builder credibility with technical founders / engineering teams”