AI for mobile apps, built by people who build mobile apps
On-device and cloud AI features in native iOS and Android apps: semantic search, summaries, voice, camera understanding and personalisation, with a router that chooses the right place to run each one.
A router inside the app decides where each request runs
Fast, private, offline tasks stay on the phone. Heavy reasoning goes to a cloud model through your backend. The decision is made per feature and, where it matters, per request based on network, battery and sensitivity.
The hybrid pattern we use in most apps. Keys and prompts live in the backend; the device holds a small model and a cache; flags let you change the route after launch.
On-device versus cloud
The trade-offs, feature by feature
Concern
On-device model
Cloud model via your backend
What we usually choose
Latency
Immediate, no network round trip
Depends on network and model; streaming helps
On-device for typing-speed interactions
Privacy
Data never leaves the phone
Data is sent; redaction and provider terms matter
On-device for personal or health data where a small model can do the job
Offline use
Works anywhere
Needs connectivity
On-device for field, travel and event apps
Reasoning and context
Small models; short context; narrower tasks
Frontier models; long context; tools and retrieval
Cloud for reasoning, summarising, answering from your content
Cost per use
Zero after download
Per-token cost; caching and routing reduce it
On-device for high-frequency, low-value calls
Updates
Model ships with the app or as a download
Change prompt or model server-side at any time
Cloud where behaviour will change often
Battery and thermal
Measurable load; needs scheduling and quantisation
Semantic search over the user's own content or your catalogue, with on-device embeddings for private data and cloud retrieval for shared content.
Summaries and drafting
Summarise threads, notes, documents and sessions; draft replies in the user's voice. Cloud route, with caching so the same content is never summarised twice.
Voice input and dictation
On-device transcription for speed and privacy, cloud for accuracy on long recordings. Built on our speech-to-text work.
Camera understanding
Scan documents, badges, labels and receipts; describe scenes; read text. See computer vision for the pipeline behind it.
Personalisation and recommendations
Rank content, sessions or products for each user. Our own Eventify event apps use this for agenda suggestions and attendee matchmaking; see event apps.
In-app assistant
A grounded helper that knows the screen the user is on and the plan they are on. The mobile face of our chatbot builds.
Process
Four to eight weeks per feature
Choose the route
Feature goals, data sensitivity, offline needs, device mix. On-device, cloud or hybrid decided with evidence.
Build the backend route
Gateway, prompts, caching, evals and cost tracking as in our LLM app development work.
Build the device side
Model selection and quantisation, router, cache, flags, graceful fallbacks when offline or throttled.
Test on real devices
Battery, thermal, memory, low-end devices, poor networks. Privacy labels and store notes prepared. Staged release.
What goes wrong
The three mistakes we see most: an API key embedded in the app binary, a cloud call on every keystroke with no cache, and an on-device model chosen on a flagship phone that makes mid-range devices overheat. Each is avoidable in design and painful after launch.
Should AI features run on the device or in the cloud?
Both, usually. On-device models win for latency, privacy and offline use but are limited in size and reasoning. Cloud models win for reasoning, long context and up-to-date knowledge but need a network and cost per call. We design a router that picks per feature, and often per request.
Do you build the app as well as the AI?
Yes. Teksmobile has delivered more than 1500 apps since 2006 in Swift, Kotlin and React Native. The AI features are built by the same team that builds the rest of the app, which avoids the usual seam between an AI vendor and a mobile vendor. See custom software development.
How do you keep an AI feature from draining the battery?
By budgeting it like any other resource: small quantised models, work scheduled when the device is charging where possible, caching of results, and the cloud route for anything heavy. We measure thermal and battery impact on real devices before release.
What about app store review and privacy labels?
AI features change your privacy disclosures: what data leaves the device, which third parties process it and why. We document the data flows and prepare the privacy labels and review notes as part of the release.
How long does it take to add an AI feature to an existing app?
A single feature such as semantic search or summarisation typically takes four to eight weeks including the backend route, evals and device testing. Voice or camera features that need on-device models take longer because of model selection and quantisation.
Talk to us about your AI project
Tell us what you are trying to do, which systems are involved and what is in the way. We reply within one business day with an opinion and a suggested next step.