Week 5 — Ship
5 lessons · project on day 30
- Open
Day 25
Monitoring
Tracing a generation end to end, logging responsibly, and catching cost, latency, and quality drift before users do.
- Open
Day 26
Scaling
Caching, batching, queueing, backoff, and failover — the engineering that keeps an AI product working (and affordable) under real load.
- Open
Day 27
Portfolio
What an AI-engineering portfolio must prove, and how to write it up so a hiring engineer can evaluate it in 90 seconds.
Day 28
Interview
Day 29
Job Ready
Day 30 project
AI SaaS Capstone
Take one of the previous four projects and make it a product: usage metering per user, a hard quota that returns a real error instead of a surprise bill, structured logs on every model call, and a deploy that someone else can sign into. This is the whole 30 days compressed into one shippable artifact.
Stack
- Next.js
- Postgres
- Auth
- Rate limiting
- Structured logging
- Vercel
The hard parts
- Metering tokens accurately when responses stream and can be aborted halfway through.
- Enforcing a quota atomically — two concurrent requests both reading 'user is at 99 of 100' will both be allowed through without a transaction or an atomic increment.
- Keeping cold-start latency tolerable once auth, database, and model calls are all on the critical path.
Why it belongs in your portfolio
A live URL a stranger can use, with quotas that hold under abuse, is worth more than five half-finished demos. It is also the only project here that proves you thought about cost.