NyotaNow: what broke while building invitations for Indian families

Every birthday invite I get is a JPG forwarded on WhatsApp, and every one of them is followed by the same three messages: “aa rahe ho?”, “location bhejo”, “kitne log aaoge?” The host ends up keeping a headcount in their head, and the image can’t answer any of those questions. It can’t open a map, it can’t collect a reply, it can’t tell you who’s coming.

NyotaNow is built on one idea: an invitation should be a link, not a picture. The host fills in a few details, picks a design, and shares one link. Guests get a live countdown, a Google Maps button, add-to-calendar, and a one-tap RSVP. The host gets a private page with the guest list. Hindi and English are both first-class, with real Devanagari typography rather than a translation bolted on at the end.

It went from idea to production in four evenings, and it’s live now. What follows is the parts that didn’t work the first time, because those are the parts worth writing down.

One card, three places it has to look identical

The invitation card is rendered in three different contexts: the live preview while the host types, the public invite page, and a 1080×1350 PNG that guests can download for WhatsApp status. Keeping three copies of that layout in sync would have been a maintenance trap, so the card is a single component sized entirely in container-query units (cqw). Every font size, padding and border is a percentage of the card’s own width, so the same markup scales from a 320px thumbnail to a 1080px export with no duplicated styles.

The bug: a rounding rule that scrambled Hindi titles

Downloading the card worked perfectly in English and broke in Hindi: the title collapsed onto one line and the text below overlapped it. My first assumption was fonts, so I checked whether the Devanagari font was embedded (it was), then whether the export picked a different fallback font (it didn’t — I changed the font stack to rule it out and the bug survived).

The actual cause was in the export library’s style-copying code: it rounds every font size down to floor(px) - 0.1. On a 340px card, a 34.4px title becomes 33.9px. That’s invisible, but my title sat right at a line break, so a fraction smaller made it fit on one line — while the copied element height stayed two lines tall, which pushed everything below out of place.

The fix was structural rather than cosmetic. Every text line now stretches to the card’s full width and centres its text, instead of shrink-wrapping its own content, so line breaks depend on the card’s width rather than on exact glyph widths. The export also renders from an off-screen copy at the full 1080px, so text is rasterised at native size instead of being scaled up. The lesson I keep relearning: when it looks like a font problem, check the numbers first.

Designing for Devanagari, not translating into it

A few details that a string-translation approach would have missed. English headings use wide letter-spacing, which looks elegant; applying the same to Hindi breaks the shirorekha, the horizontal line that joins letters into words, so Hindi text gets none. Hindi cards close with स्नेहाकांक्षी or निवेदक, the words actually printed on Indian invitation cards, instead of a literal “With love”. And the link preview image that WhatsApp generates can’t shape Devanagari at all — the renderer behind it does no complex text shaping — so for Hindi invites the preview falls back to English wording while the invite page and the downloaded image stay fully Hindi. Not perfect, but an honest trade-off rather than broken glyphs in every chat.

No accounts, without being insecure

Nobody signs up to invite people to a birthday, so there are no accounts at all. Each invite issues a random 24-character key that lives in the host’s private link. Only a SHA-256 hash of it is stored, and it’s compared in constant time, so response timing leaks nothing. A wrong key returns exactly the same 404 as an invite that doesn’t exist, which means links can’t be probed for valid ones.

Migrations that survive serverless cold starts

Schema migrations run automatically on the first database connection, which is convenient until a deploy cold-starts several instances in the same second and they all try to create the same tables. A Postgres advisory lock makes them take turns: the first instance takes it, the rest wait and then find nothing left to do. The transaction-scoped variant releases itself even if a process dies, so the lock can never get stuck.

The same Drizzle queries run against two different databases: an in-process PGlite instance locally, so the project needs no Docker and no setup, and Neon Postgres in production. Nothing above the data layer knows which one is live.

Shipping it, and one deploy trap

Vercel’s database integration pre-fills the environment variable prefix as STORAGE, which would have created STORAGE_URL. My code reads DATABASE_URL. The build would have passed, the homepage would have loaded, and the site would have failed on the very first invite. Defaults are worth reading.

The first deploy also ran in Washington, D.C. by default, with the database beside it — for a product whose users are all in India. Since a Neon database’s region is fixed at creation, the fix was to recreate it in Singapore while it was still empty and move the functions there too. Requests now enter Vercel’s network in Mumbai and run in Singapore next to the database; a warm invite page responds in about 0.3s from India.

Measuring the thing that matters

Every invite footer says “Made with NyotaNow, create your own”. One birthday invite is opened by 30 to 80 guests, and some of them have an event coming up — that loop is the entire growth plan, so it needs a number. Invites record whether the host arrived through another invite’s footer, and a private stats page shows that share alongside invites created, RSVPs and replies per invite. If the share climbs, the product spreads on its own; if it doesn’t, no amount of posting will save it.

What’s next

Rate limiting, a test suite and CI, error monitoring, and eventually a small paid tier to remove the branding. But first: real invites, from real hosts, to real guests. Everything else is guesswork until then.

Try it at nyotanow.vercel.app — it’s free and needs no sign-up. The code is at github.com/shane-Coder/nyotanow.