A true story
So I built a small team of AI agents — using OpenAI, Sarvam, Cohere and Cloudflare — to do exactly that.
Let me show you what I built — and how I built it — with the tools I just named.
Chapter 1
Before the day has decided what it wants from her, my mom picks a good-morning image and sends it. A sun, a flower, a deity, a line in Hindi. These are the actual ones — watermarks and all.
Nobodymakestheseforher.SoIdid—well,Ibuiltateamthatdoes.
Good morning beta 🌞 Have a blessed day.
Exactly as received. Watermarks and all.
Chapter 2
One agent on stage at a time — what it actually does, and what it hands to the next. Every brief, shot and picture below is real.
Generates the piece of content. That's it — no opinions, no edits.
A series of strict gates and hard grades. Decides what reaches her feed — and what gets trashed.
…and every agent is told what was trashed. The system learns from its own mistakes.
This runs on its own, end to end, on Cloudflare. The only thing a human does here is this button.
Meet the whole constellation →Meanwhile
While you scrolled, the relay kept going. This is a slice of a single morning — every faith, every language, every festival within reach.
✦ zero humans involved
And every single one of them now faces the judge ↓
Chapter 3
Everything you just saw was made by AI. So who checks it? Another AI — hired only to judge. Every greeting gets a scorecard: seven metrics, a composite, a verdict. The dangerous ones are vetoed in code — and a human can always overrule.
Pretty images that don't forward are worthless; wrong images that do forward are worse.
the judge, verbatim“The image accurately depicts a Raksha Bandhan scene with appropriate motifs and text. There are no iconographic errors or cultural insensitivities present.”
the judge, verbatim“The silhouette in the background is not clearly identifiable as Mahabali, it could be mistaken for a Buddhist stupa.”
the judge, verbatim“Text is missing the intended suffix 'हनुमान जी के आशीर्वाद से।'”
the judge, verbatim“shadowy, frightening imagery”
scroll back — the stamp un-slams
The verdict, at scale
The judge doesn't review one card — it sweeps the whole shelf. Watch one morning get judged:
The kept ones are already on their way to her ↓
✓ kept ⚑ flagged ✓ kept ✓ kept ✕ trashed ✓ kept ✓ kept ⚑ flagged ✓ kept ✓ kept ✕ trashed ✓ kept ✓ kept ⚑ flagged ✓ kept The app
Everything the agents make — and the judge keeps — lands in an app called Namaste Ji: feeds she can scroll to find what she likes, personalized, tuned for daily forwards, culturally accurate and aware of what day it is.
Chapter 4
The feed knows it's Friday — Mata Rani's day — knows her language, knows what she loves. And every card in it passed the judge. She scrolls… stops… forwards.
or tap ⤴ yourself
Good morning beta 🌞
Thestoryendswhereitbegan.Exceptnowthecardwasmadeforher.
Play with the feed's brain →Chapter 5
I built all of this alone. Well — me and a team of coding agents. Claude Code wrote most of what you're looking at, this site included, with me directing. It was the most fun I've had building anything.
And it's not a pet project. Queues with dead-letter lanes, RBAC from day one, infrastructure as code, versioned prompts, eval pipelines — a stack you could take to production and do insane things with. So I boxed it up.
The stack, in orbit hover to hold · click to visit
Clone https://github.com/dashayushman/karkhana and read AGENTS.md. Put my Cloudflare API token in .env, run `npm run bootstrap`, and verify the daily-card example deploys. Then build me ______ on the same architecture.
and underneath, all Cloudflare:
Chapter 6
That's the story. From here — watch it work under the hood, or learn to build one yourself.
