Kata Jaam

The problem

Kathmandu has a real nightlife, a running scene, concerts, screenings and festival events every week, and no single place that lists them. Plans lived in Instagram stories that vanish after a day. Organizers had reach only as far as their own followers.

I wanted the thing I kept wishing existed: open one app on a Friday and know what is on tonight. I also wanted to find out whether one engineer with AI agents could ship and run a two-sided marketplace alone.

What I built

A Flutter app on both stores, a public website that Google indexes deeply, an organizer portal, an admin panel and a set of Cloud Functions behind them all. The first version shipped to the stores two weeks after the first commit.

• The app: home spine (tonight, weekend, scenes), a Google Maps view with named pins, event pages with Interested, calendar export, directions and a chat, crews for going with strangers, push notifications that are targeted rather than broadcast.
• The web: every event, venue, scene and festival has its own page with JSON-LD, a static map and canonical rules for recurring series. This is where the users come from.
• Organizers: claim your listing, run a registration form, see responses, message every registrant from one screen, buy a placement. Native ticketing is built and waits on the payment gateway.
• Admin: a decision desk that shows what needs a human now, and nothing else. Everything the agents do lands here with an approve or reject.

The operation runs on agents

The part I am most proud of is not the app. It is that the company runs with about fifteen minutes of my attention a day. A set of Claude Code agents, each with a written brief, work the Instagram inbox and the email inbox, walk Instagram to find new events and import them through a CLI, review the catalogue for errors, draft the social posts, run a weekly SEO loop and a weekly retention loop, and write the day's report.

Every outward action files on an approvals desk first. The agents never publish, never message an organizer, never approve a listing without a human click. The hard limits are written down; the judgement is theirs. A nightly shift takes the bugs the day found and builds the fixes in an isolated worktree so I wake up to diffs.

I wrote more about how this works on the agents page.

Decisions that mattered

• Search first, not social first. Instagram gave reach for a week at a time; a page that ranks gives it every day. Most signups say they found us on Google.
• Interested is the one save signal. A separate bookmark split the data and taught nothing; it was merged and the counter became meaningful.
• Notify, never ask permission. Organizers are listed, credited and told; a takedown is one click. Asking first meant no catalogue.
• Keep the Admin SDK on REST on Vercel. The default gRPC transport wedged on reused instances and turned a cache into three million Firestore reads a day. One flag fixed the bill.
• One tombstone for every deleted event, so a URL Google has already indexed never 404s into a lost ranking.

What changed after launch

Users went from 0 to 2,451 in eight months on no paid acquisition. The site holds first position for the head term and ranks for hundreds of long-tail event queries. Weekly planners, the north-star metric, grew 11 percent week on week in the most recent read. The first organizers paid for placements in October 2026.

The honest part: week-four retention sits under 10 percent, because most people plan one night and come back the next time they need a plan. The push rework and the scene follows exist to change that, and the number is still being read every Thursday.