The whole system, on one page.

Namaste Ji is a real, running product: five AI agents research Indian festivals, decide what to make, generate devotional greeting images, quality-check them, and feed them to an app — while one human supervises from a dashboard. It runs entirely on Cloudflare. This track walks through how, top-down: one picture, then eight zooms into it.

Why this might matter to you.

Almost nothing here is specific to greeting cards. Swap the five agents' jobs and the same chassis builds any app where AI agents produce the product and a human stays in charge: research agents feeding a newsletter, listing agents feeding a marketplace, support agents feeding a knowledge base. The pattern is always the same four pieces — agents that make, a store they share, a human gate, and a serving path that stays dumb and fast. The last zoom hands you this chassis as an empty repo.

Six words you'll keep meeting.

Everything else is explained where it appears — these six come up everywhere, so here they are once, in plain words.

Agent: A small program with one job. It runs on a schedule, thinks with an AI model, uses tools, and writes its results to a database. Not a chatbot — nobody talks to it.
Brief → shot: A brief is an idea for a batch of content ("Diwali morning cards, Hindi, warm"). A shot is one exact picture to make, with the words for it fully written out.
The catalog: The one database of record. Rows in D1 (SQLite), the images themselves in R2 (file storage), and a searchable map of what each image looks like in Vectorize.
Embedding: A list of numbers describing what a picture looks like, so "these two are similar" becomes something you can compute. Made once per image, reused everywhere.
The console: The private dashboard where a human watches every run, approves work, and can pause any agent. A read-only copy is open to you — no login.
Slate: A pre-computed list of ~50 cards waiting in storage for one person's next morning. Built overnight, served instantly.

The picture.

Content flows top to bottom. The agents (top left) make greetings into the catalog (the middle band); the feed machinery (bottom left) picks from the catalog what each person sees in the app (bottom right). The human's dashboard (top right) watches and gates all of it, and the strip at the bottom is the infrastructure everything runs on. Click any region to zoom in.

Two rules the whole design follows.

AI decides slowly, code serves fast

Models do their thinking in scheduled background work — planning, generating, judging, personalising. The moment a person opens the app, everything they touch is plain deterministic code. That keeps the product fast, cheap, and debuggable.

A human can stop anything

Every agent runs inside a gate a person can flip — automatic, manual, or paused — from the dashboard, without a deploy. Autonomy is a setting, not a leap of faith.

The zooms.

Eight, in order — each opens one region of the picture above. Stop at any point and you still hold a correct picture; go all the way and the last zoom is a repo you can build your own system on.