CarrySo
Menu
Service

AI app development: AI-powered mobile apps built with Flutter

We build AI-powered mobile apps in two ways: on-device models that run privately on the phone, which is how our own apps remove photo backgrounds and analyse media, and LLM-backed features that call hosted models through a secure backend. Both ship inside Flutter apps for Android and iOS with Firebase behind them.

On-device AI
Image segmentation and media analysis running offline in our published apps
Hosted models
Anthropic, OpenAI and Google model APIs called from Cloud Functions, never from the app
App
Flutter for Android and iOS
Backend
Firebase Auth, Firestore, Cloud Functions, usage limits and cost controls
Discuss your app See our work

Two kinds of AI features, and when each fits

On-device versus hosted AI in mobile apps
On-device modelsHosted LLM / vision APIs
RunsOn the phone, offlineOn a server through your backend
PrivacyData never leaves the deviceData is sent to the model provider; needs consent and policy work
CostNo per-request cost; larger app sizePer-token or per-image cost; needs limits
Best forPhoto and video tools, background removal, classification, OCRChat assistants, summarisation, generation, reasoning over user data
Our evidenceBG Remover AI (offline segmentation), MetaClean (media analysis)Our internal Growth OS tooling calls hosted models from Cloud Functions

AI features we build into apps

  • Offline photo and video tools: background removal, enhancement, classification, metadata analysis
  • Chat and assistant features backed by hosted models, with conversation history and guardrails
  • Summarisation, extraction and generation over the user's own content
  • Search and recommendations using embeddings
  • Voice and text input with on-device speech recognition where the platform supports it

How an LLM-powered mobile app is architected

The app never holds a model API key. Requests go to a Cloud Function that checks the user's session and entitlements, applies rate and cost limits, adds the system prompt and any retrieved context, calls the model, and returns a response the app can render. Prompts, model choice and limits live in configuration, so they can change without a store release.

  • Authentication and per-user quotas before any model call
  • Cost controls: token limits, daily caps and a cheaper fallback model
  • Streaming responses for chat-style features
  • Logging without storing sensitive user content longer than needed
  • Evaluation prompts and test cases so model changes do not silently break features

Store rules and responsible use

Both stores require clear disclosure of AI-generated content, working reporting for objectionable output, and privacy declarations that match what the model provider receives. We design the consent screen, the Data safety and privacy label answers, and the moderation path as part of the feature, not after review rejects it.

AI app project process

  1. 1. Discovery and scope

    We go through who the app is for, the one job it must do well, and what can wait. You get a written scope with the screens, features, integrations and release target.

  2. 2. UX flows and interface design

    We map the main user journeys before writing code, then design screens that follow Material guidelines on Android and feel right on each platform.

  3. 3. Build in short iterations

    The app is built in Flutter in small increments, so you can install test builds on a real phone early and change direction cheaply.

  4. 4. Testing

    Business logic, parsers and data layers get automated unit tests, and release builds are checked on physical devices before they reach users.

  5. 5. Store release

    We prepare the store listing, screenshots, privacy policy, Data safety form and release build, then handle review and the staged rollout.

  6. 6. Maintain and improve

    After launch we watch crashes, reviews and release age, keep SDKs and target API levels current, and ship improvements.

Evidence: ai app development in our own apps

These are CarrySo's own products, published on Google Play under our developer account. Figures are taken from Google Play when this site was built; each card links to the case study.

Frequently asked questions

Can a Flutter app run AI models on the device?

Yes. Our published apps run image segmentation and media analysis entirely on the device with Flutter, in background isolates so the interface stays smooth. Larger language models still belong on a server.

Which AI providers do you work with?

We integrate Anthropic, OpenAI and Google model APIs, and OpenAI-compatible endpoints, always from a backend so keys and costs stay under your control.

How do you keep AI costs under control?

Per-user quotas, token limits, daily budgets and a cheaper fallback model, all enforced in Cloud Functions and adjustable without an app update.

Mobile app development

Mobile app development from idea to app store: Flutter apps for Android and iOS, SaaS mobile apps, MVPs, UI/UX, backend APIs, store deployment and maintenance.

Flutter app development

Flutter app development for Android and cross-platform apps: architecture, state management, Firebase, on-device processing, testing and store release.

Mobile app backend and API development

Backend and API development for mobile apps: Firebase (Auth, Firestore, Cloud Functions, FCM), REST API integration, security rules, admin panels and remote config.

SaaS mobile app development

Mobile apps for SaaS businesses and mobile-first SaaS products in Flutter: subscriptions, auth, onboarding, dashboards, payments, push, APIs and admin panels.

Tell us what you want to build. We'll reply by email with questions, and then with a proposed scope and estimate.

Start a project Email carrysoofficial@gmail.com