Case Study
Shipping a multi-modal AI translation app to Google Play
Ferhengy is an AI translation app built for a client in Türkiye and published on Google Play. It supports 85+ languages with a deliberate focus on Kurdish, covering both Kurmanji and Sorani. Users can translate by typing, speaking, pointing a camera at printed text, or uploading a PDF, and hear any translation read aloud. Every feature works without creating an account. ManiWebDev built the React Native app, the Node.js backend, the AI service integration layer, and the production infrastructure it runs on.
The challenge
Kurdish speakers are poorly served by mainstream translation tools. Where major languages get voice input, camera translation and natural-sounding speech output, Kurdish often gets a text box, if it appears at all. Kurmanji and Sorani are spoken by tens of millions of people, and the tooling does not reflect that.
The brief was to build a translation app where Kurdish is treated as a primary language, not a checkbox.
That single requirement drove almost every technical decision. No single translation provider handles Kurdish speech, general translation, and 84 other languages equally well. Meeting the brief meant building an integration layer that routes each language and each input type to whichever service actually handles it properly, and hides all of that from the person using the app.
What we built
- 01
Five Ways to Translate, One App
Text, voice, text-to-speech, camera, and PDF. Each input method is a separate technical pipeline, and all five had to feel like one product.
Typing uses automatic language detection, so users never pick a source language. Voice records, transcribes, translates, and speaks back. Camera reads printed text from a live view or the gallery. PDF upload extracts text and returns it translated. The complexity sits underneath; the interface stays a single screen.
- 02
Kurdish as a First-Class Language
Ferhengy supports Kurmanji and Sorani with the same depth as major world languages, not as an afterthought at the bottom of a dropdown.
Kurdish is underserved by mainstream translation apps. Getting it to parity meant sourcing and integrating speech services that handle it properly, rather than accepting whatever a single general-purpose provider offered. The product exists because of this gap.
- 03
Privacy-First Architecture
Every feature works without an account. Translation history stays on the device. An account is optional and exists only to sync bookmarks across devices.
This was a design decision with real engineering consequences. Local-first storage, no forced signup wall, and an auth system that is genuinely optional rather than delayed. Most apps take the opposite route because accounts are easier to build around.
- 04
Production Infrastructure
A Node.js backend on AWS EC2 running under PM2, behind HTTPS, with rate limiting, security headers, and structured logging.
Multiple third-party AI services sit behind a single backend layer. That means one place to handle failures, retries, rate limits, and cost control, and it means any provider can be swapped without touching the mobile app.
Decisions that shaped the build
Integrate, do not reinvent. Every AI capability in Ferhengy comes from an established service rather than a model built from scratch. That was the right call: faster to launch, cheaper to run, and more reliable than anything a solo build could have matched. The engineering value is in the routing and integration layer, not in owning a model.
One backend in front of every provider. The app never talks to an AI service directly. API keys stay on the server, cost and rate limits are controlled in one place, and a provider can be replaced without shipping a new app version.
No signup wall. The single biggest drop-off point in a consumer app is the account screen on first open. Ferhengy has none. Everything works immediately, data stays local, and an account is offered later only for syncing bookmarks.
Automatic language detection everywhere. Asking a user to select a source language before translating is a small friction repeated hundreds of times. The app detects it instead.
Hardened from the start. HTTPS, security headers, rate limiting on API routes, and structured logging were built in rather than added after launch. An app that proxies paid AI calls is a target for abuse from day one.
The stack
- Mobile
- React Native (Android)
- Backend
- Node.js, Express
- Database
- MongoDB Atlas
- Auth
- JWT, Google OAuth, OTP
- AI services
- Third-party translation and speech APIs
- Infrastructure
- AWS EC2, PM2, HTTPS
- Hardening
- Rate limiting, security headers, structured logging
The app is live on Google Play and maintained through ongoing releases. View Ferhengy on Google Play.
Common questions
What is involved in building a multi-modal AI app?
Each input type is its own problem. Voice needs audio capture, format conversion, and transcription. Camera needs image handling and text recognition. PDF needs document parsing and extraction. The engineering work is less about any single AI service and more about making four different pipelines return results that feel identical to the user.
Do you build with existing AI services or custom models?
Existing services, chosen carefully. For most products, a well-integrated commercial API beats a custom model on cost, reliability, and time to launch. The value is in the integration layer: routing each language and input type to the provider that handles it best, handling failures gracefully, and keeping providers swappable.
Why route AI calls through your own backend instead of calling them from the app?
Three reasons. API keys never ship inside the mobile app. Rate limiting and cost control happen in one place. And providers can be changed without publishing a new app version and waiting for users to update.
Can an app work properly without forcing users to sign up?
Yes, and Ferhengy does. Every core feature runs without an account and data stays on the device. An account is optional and only adds cross-device sync. It takes more design effort than a signup wall, and it removes the biggest reason people abandon an app on first open.
Do you build mobile apps as well as web systems?
Yes. Ferhengy is React Native on Android, published on Google Play and maintained through ongoing releases. The same backend and AI integration work applies whether the front end is a website or a mobile app.
Have an AI product idea that needs building?
Whether it is a mobile app, a client-facing system, or an AI layer on top of what you already run, the hard part is usually the integration rather than the AI. Tell me what you are trying to build and I’ll tell you honestly what it takes.
Start a conversation