A wedding invitation that speaks ten languages

What was the problem?

A wedding spanning several events and a guest list across multiple languages and locations, where a conventional invitation — paper or a static image — cannot carry a timeline, handle a live stream, or let guests who could not travel take part at all.

What made it hard

What we built

A fully custom digital wedding invitation: a countdown to the day, an AI Wishbook turning guest wishes into personalised calligraphy and artwork, multilingual support across English, Hindi, Telugu, Marathi, Tamil, Kannada, Gujarati and Bengali, a full event timeline covering Mehendi, Bridal Day, Groom's Day, Wedding and Reception, live streaming, a guest-shared media gallery and a shareable invitation with WhatsApp integration.

What it does

• AI wish enhancement across ten languages with five tones

• Three-tier language detection with directive stripping

• Script, echo and length validation on every AI response

• Fifty hand-written fallback wishes

• Two-stage AI artwork generation in three styles

• Cost-controlled Daily Love Note generator

• No-login guest photo and video upload

• Device-fingerprint guest identity with upload attribution

• Event-sorted media gallery with likes and full-screen viewing

• Integrated live streaming

• Eleven Firestore-controlled page sections with instant visibility toggles

• Admin moderation with CSV and JSON export

• Shareable invitation image, calendar integration and email reminders

Engineering decisions

Choices that could have gone another way, and what each one bought.

1. 1

   Language resolved three ways before a single token is generated

   The system reads a written directive in the guest's text first — including one written in the target script itself — then an explicit dropdown selection, then falls back to Unicode block detection of whatever script the guest typed in. The directive is then stripped from the text so the model cannot echo the instruction back.

   Why it matters. An elder who types a blessing in Telugu script has already told the system everything it needs to know. Asking them to also find a dropdown is the kind of friction that stops people using the thing at all.

2. 2

   The AI's output is checked before anyone sees it

   Each of five tones maps to its own temperature and style hint — 0.9 for funny, 0.55 for blessing. Every response is then validated for correct script, tested for echo against the guest's own input, capped at fifty words and has the couple's names injected. Anything that fails falls through to fifty hand-written wishes covering ten languages and five tones.

   Why it matters. Generation is the easy half. The guest never sees a wrong-script response, a parroted prompt or an error message, because every failure mode has somewhere to land.

3. 3

   A language model used as a prompt engineer

   Artwork generation runs in two stages: Gemini 2.5 Pro takes the guest's rough idea plus a style guide — Cartoon, Ghibli or Royal — and writes a detailed South Indian wedding image brief, which Imagen 4.0 then renders, behind a fallback model, two retries and thirty-second timeouts.

   Why it matters. A one-line guest prompt sent straight to an image model produces generic output. Letting a language model do the describing first is what makes the results worth showing.

4. 4

   Knowing who uploaded what, without anyone signing up

   A device fingerprint built from host, user agent and platform is persisted to local storage and a one-year cookie, then synced to a Firestore guest profile holding a name and phone number. Each upload upserts that record and increments the guest's upload count.

   Why it matters. Asking a wedding guest to create an account is how you get no photographs. This gives the couple full attribution while asking the guest for nothing.

5. 5

   A CMS the couple could actually use

   Eleven page sections live in Firestore, fetched on load and gated behind a visibility check on the page. The couple toggles a section in the admin panel and it disappears from the live site instantly — and if the backend is unreachable, sensible defaults keep the site rendering.

   Why it matters. The alternative is calling a developer at eleven at night the day before the wedding.

6. 6

   Cost control designed into a feature

   The Daily Love Note calls Gemini on only every third refresh; the rest of the time it draws from thirty-nine curated quotes, each attributed to a real named family member.

   Why it matters. It caps API spend and guarantees a quality floor in the same decision — the fallback content is arguably better than the generated content.

7. 7

   No single point of failure, because there is only one wedding day

   Every AI route has a hand-written fallback, section and settings configuration fall back to defaults, analytics is fire-and-forget, media deletion is idempotent, and every external call carries a timeout.

   Why it matters. A wedding website gets exactly one chance to work. If an external service fails mid-ceremony, the site keeps running.

Technical detail

Collapsed by default. Open whichever part you are evaluating.

Before and after

The build, by the numbers

What was built, at what scale. These describe the system's size, not its business results.

languages supported

10languages supportedWish enhancement across ten Indian languages with script detection

hand-written fallback wishes

50hand-written fallback wishesTen languages across five tones, so a guest never sees an AI error

runtime-configurable page sections

11runtime-configurable page sectionsToggled by the couple with no code change or redeploy

AI models in the pipeline

3AI models in the pipelineGemini 2.0 Flash, Gemini 2.5 Pro and Imagen 4.0, plus a fallback image model

What changed as a result?

A modern, tech-forward wedding experience that goes far beyond a traditional invite: guests participate, share moments and stay connected through one custom-branded digital touchpoint.

• Eight languages supported
• Live streaming for guests unable to travel
• AI Wishbook turning guest messages into personalised artwork
• Guest-shared media gallery, so the record built itself

See it running