Hi. I run engineering at Utsav, where devotees across India book a puja from their phone, a temple performs it, and the video (and often the prasad) finds its way back to them. Small team, big surface: two mobile apps, a website, an admin platform, nine repos, shipping in Tamil, Telugu, Kannada, Hindi and English.

I like systems that keep working at 3 am, and tooling that gives the team an afternoon back.

• Booking and payments that survive three gateways, retries, refunds, and the occasional webhook that arrives twice.
• Ops tooling for a team that talks to customers on WhatsApp all day: call trackers, lead routing, schedule dashboards.
• A video pipeline that turns raw temple footage into a finished puja video without anyone sitting at a timeline.
• An agent for every department. More on that below.

Not demos. Things the team actually runs every week.

• One Claude Code agent per department: marketing, SEO, BI, support, ads, push. Each has its own context, prompts and skills, versioned in a repo like any other code.
• The monthly engineering report writes itself from the git history of every repo plus the team's daily log, then renders to a five-page PDF for the all-hands.
• The new-developer onboarding kit regenerates from the live code: ER diagrams checked against the models, ports, env names, versions. It is never stale on someone's first day.
• Before anyone touches an API shape or a database column, an agent traces the blast radius across all the repos and lists every consumer.
• In the product: puja and event copy drafted and translated into five languages, and support chat on WhatsApp and social, on a model-agnostic service. A human approves every word a customer sees.

The rule is simple. AI drafts, people decide, and everything it touches shows up in a diff.

• PRD first. But the PRD is written from what the code does today, not from what we wish it did.
• Docs live in the repo as markdown, next to the code they describe. If it only exists in a chat thread, it doesn't exist.
• One repo, one branch, one PR. Anything that crosses repos gets its consumers listed before a line is written.
• Every user-facing string is a five-language change. No "we'll translate it later".
• Dense data, calm tables. Weight over colour.
• I'd rather read the log than the slide.

Both derived from MIT-licensed projects and adapted for my own needs: image-stories, from react-insta-stories, and i18n-next-export, from next-export-i18n.

LinkedIn · X · utsavapp.in · the company org, @applex-in

_{The green squares below are mostly private repos under @applex-in.}