The conversation around AI in mobile has moved fast. A year ago, integrating a Large Language Model into an iOS or Android app meant complex API calls, unpredictable latency and significant infrastructure overhead. Today, the tooling is mature, the APIs are stable, and on-device inference is accessible on consumer hardware. Here is what is actually possible right now — and how we implement it.
Cloud LLMs in Mobile Apps
The most straightforward AI integration is calling a cloud LLM from your app. The main providers — OpenAI (GPT-4o), Anthropic (Claude) and Google (Gemini) — all offer REST APIs with streaming support. On iOS, Swift’s async/await and URLSession make it easy to stream tokens directly into SwiftUI views, giving users the familiar “typing” effect. On Android, Kotlin Coroutines and Flow handle the same pattern cleanly.
Practical use cases that work well today:
- In-app assistants and chatbots with full conversation history
- Document analysis (upload a PDF, ask questions about it)
- Smart form filling and content generation
- Voice-to-text with Whisper, combined with LLM processing of the transcript
- Summarization of long content (articles, reports, emails)
Retrieval-Augmented Generation (RAG) on Mobile
RAG lets an LLM answer questions using your own data rather than just its training data. In a mobile context, this usually means the app sends a query, a backend retrieves relevant chunks from a vector database, and the LLM generates an answer grounded in that retrieved content. The result is an assistant that knows about your product, your documentation, or your user’s personal data — without hallucinating answers it doesn’t know.
We implement RAG pipelines with Pinecone, Weaviate and pgvector, connected to mobile apps via REST APIs we build on top of them.
On-Device AI: Core ML and TensorFlow Lite
Not every AI feature needs a server. Apple’s Core ML and Google’s TensorFlow Lite / ML Kit let you run models directly on device — with zero latency, no API costs, full privacy and offline capability.
What works well on-device today:
- Image classification and object detection (MobileNet, EfficientDet)
- Natural language processing for short texts (sentiment, intent classification)
- Face and body pose detection (Vision framework on iOS, ML Kit on Android)
- OCR and document scanning (Vision, ML Kit)
- Small language models (Phi-3 Mini, Gemma 2B) — surprisingly capable for specific tasks
Hybrid Approaches
The most effective AI apps use both layers. A typical pattern we implement: a fast on-device model handles common, latency-sensitive tasks (intent classification, quick suggestions), while a cloud LLM handles complex queries that require reasoning or large context windows. The user gets instant responses for 80% of interactions and waits only when they genuinely need the heavier model.
Privacy Considerations
Any AI integration that sends user data to a third-party API raises privacy questions. We address this from the start: request anonymization, no-log API configurations (OpenAI and Anthropic both offer these), on-device alternatives where feasible, and clear privacy policy language that covers AI data processing. GDPR compliance is not optional for European clients.
Getting Started
The best AI integrations are ones that solve a real user problem — not features added to check a box. We always start with the use case: what does the user need to do faster or better? Then we choose the right AI layer (on-device, cloud, or hybrid), integrate it natively, and measure whether it actually helps.
If you have an iOS or Android app and want to explore what AI could do for it, get in touch with us.
No responses yet