Featured case study
Building an AI-Powered Wardrobe & Style Intelligence App
An AI-powered wardrobe intelligence app that helps people organize their clothing and understand their personal style.
Read the case study
On-Device AI App Development
AI features that run on the phone itself: faster, private, and working offline.
Stack we work in
- ONNX Runtime
- PyTorch
- TensorFlow
- Hugging Face
- Swift
- SwiftUI
- Kotlin
- Jetpack Compose
- Projects delivered
- 680+
- Clients served
- 310+
- Client retention
- 98%
- Certified engineers
- 20+
- Building since
- 2018
Figures as of Oct 2026
Trusted by
- Aralia
- Rosa Pistrina
- Lashbo
- Micromunchi
- Payroll Ninja
- POD
- Petdx
- Yokey
- Spin Up
- Starwear
- Udlins
- Vegpal
- Wander Wings
- Wish It
- Wisl
When on-device AI makes sense
On-device AI means the AI model runs on the user's phone, tablet or watch instead of on a remote server.
Read the full overview
Features respond instantly, keep working without a connection, cost nothing per request, and personal data never has to leave the device. We add on-device AI to iOS and Android apps using the models built into the platforms, such as Apple's Foundation Models framework and Gemini Nano through ML Kit, and run custom models with Core ML, LiteRT and ONNX Runtime when you need your own. We also combine on-device and cloud models where a task needs more power.
Apps that handle sensitive data
Health, finance, legal and personal content that users do not want sent to a server.
Apps used with poor signal
Field, travel and outdoor apps that must work offline.
Features that must feel instant
Camera, typing and voice features where a round trip to the cloud is too slow.
High-volume AI features
Features used many times a day where per-request cloud fees would add up.
Outcomes
What you get
- AI features that respond instantly, even offline
- Personal data kept on the user's device
- No per-request cloud AI bill for common tasks
- Features that work on the devices your users actually own
- A clear plan for which tasks run on device and which in the cloud
What we deliver
On-device AI features we build
Smart summaries and replies
Summaries of notes, messages or documents and suggested replies, generated on the phone.
Document and receipt scanning
Capture, read and extract fields from documents without uploading them.
Learn moreVisual search and recognition
Identify products, plants, parts or defects from the camera.
Learn moreVoice notes and dictation
On-device speech recognition and transcription.
Fitness and health insights
Activity recognition and trends from sensor and wearable data.
Learn morePrivate search
Search across the user's own content with on-device embeddings.
What you receive
Feasibility report
Model test results on real devices, with a recommended approach and fallbacks.
AI features in your app
Integrated, tested and released on iOS, Android or both.
Optimized models
Converted and compressed model files with the scripts to rebuild them.
Evaluation set
Test examples and a repeatable way to measure AI quality on each release.
Built-in platform models
Use the language models already on the device.
- Apple Foundation Models framework for text generation, summaries and tool calling on iOS
- Gemini Nano through ML Kit GenAI APIs for summaries, rewriting and proofreading on Android
- Apple Intelligence features through App Intents and Writing Tools
- Image description and on-device text extraction
- Fallback to a cloud model where a device is not supported
Built with
- Foundation Models framework
- App Intents
- Gemini Nano
- ML Kit
- ML Kit GenAI APIs
Custom models on device
Run your own models efficiently.
- Model conversion and quantization for Core ML and LiteRT
- Image classification, object detection and segmentation
- Small language models run with MediaPipe LLM Inference, ExecuTorch or llama.cpp
- Speech recognition and text-to-speech on device
- Use of the Neural Engine, GPU and NPU for speed and battery life
Built with
- Core ML
- LiteRT
- MediaPipe
- MediaPipe LLM Inference
- ExecuTorch
- llama.cpp
- Llama
Vision and sensor features
Understand what the camera and sensors see.
- Document, barcode and text scanning with Apple Vision and ML Kit
- Pose, hand and face tracking with MediaPipe
- Activity and anomaly detection from motion and health sensors
- Real-time camera effects and AR overlays
Built with
- Apple Vision
- ML Kit
- MediaPipe
Hybrid on-device and cloud AI
Pick the right place for each task.
- Routing between on-device and cloud models by task, size and privacy
- On-device search over the user's own content with local embeddings
- Private Cloud Compute and cloud LLM APIs for heavier requests
- Offline queues that sync when the connection returns
Privacy, quality and rollout
Ship AI features you can trust.
- Evaluation sets to measure accuracy before and after release
- Model size, speed and battery testing on real devices
- Model updates delivered without a full app release
- Privacy design that keeps personal data on the device
Why teams choose Equitysoft for on-device AI
Since 2018 we have delivered 680+ projects for 310+ clients, and 98% of clients keep working with us.
Mobile and AI engineers together
Our mobile developers and AI engineers work as one team, so models are chosen for what the phone can actually run.
Built-in models before custom ones
We use Apple Foundation Models and Gemini Nano where they fit, and only train or convert a custom model when they do not.
Tested on real devices
We measure speed, battery, memory and app size on real phones before release, not just in a simulator.
Privacy and quality by design
We keep personal data on the device where possible and test answers against agreed examples before rollout.
Recent work from our team
AI Video Analytics CCTV Intelligence Development
Building a Secure Employee Health & Performance Management Platform
Discover Your True Personality: Fun Quizzes for Self-Insight and Growth!
Book Professional Photographers Anytime, Anywhere with POD India
Building a Real-Time Sports Data & Community Intelligence Platform
Create Stunning Art Instantly with AI HyperVerse
How we work
Discovery
We review the AI features you want, your users' devices and your privacy needs.
You receive Feature list and device targets
Feasibility
We test candidate models on real devices for accuracy, speed, size and battery use.
You receive Model test results and recommendation
Design
We design the user experience, fallbacks and the split between on-device and cloud.
You receive Architecture and UX flows
Build
Models are integrated into the app in short iterations, with an evaluation set run on every build.
You receive Working AI features in test builds
Release
Phased release with monitoring of quality, speed and crashes across device types.
You receive Feature live in the stores
Improve
We update models, prompts and thresholds as new platform models and user feedback arrive.
You receive Improvement roadmap
Tech stack
Technologies we use
AI and machine learning
Databases
Mobile
Languages
Other tools
Engagement
Ways to work with us
Fixed-scope project
On-device AI features built against an agreed scope, timeline and price.
- Written estimate up front
- Demos every iteration
Dedicated team
Mobile and machine learning engineers who work on your app full time or part time.
- Scale up or down with notice
- Works in your tools and process
Audit or pilot
A short paid review or pilot that answers the key questions before you commit to a full build.
- Clear findings and plan
- No obligation to continue
Questions, answered
It depends on whether built-in platform models are enough or a custom model is needed, how many features you want, and how many devices you support. We send a written estimate after a short call. As a guide, adding one on-device AI feature using built-in models usually costs $8,000 to $25,000, and a custom on-device model with training and app integration $25,000 to $90,000 or more (USD).
It means the AI model runs on the phone rather than on a server. Users get instant answers, features work offline, personal data stays on the device, and you do not pay a cloud AI fee for every request.
Apple's Foundation Models framework runs on devices that support Apple Intelligence, from iPhone 15 Pro onwards, with Apple Intelligence turned on. Gemini Nano runs on a growing list of recent Android phones, including Pixel and Samsung Galaxy flagships. We design a fallback for older devices, such as a smaller custom model or a cloud model.
Not for every task. On-device models are smaller, so they are best at focused tasks such as summaries, classification, extraction, rewriting and short replies. For long reasoning or broad knowledge we route the request to a cloud model, with the user's consent where needed.
Yes. We convert and compress models for Core ML on iOS and LiteRT or ONNX Runtime on Android, then test speed, accuracy and battery use on real devices before release.
It can if done carelessly. We choose model sizes and formats that use the Neural Engine, GPU or NPU, download models after install when they are large, and test battery and memory use on low-end and high-end devices.
Usually, yes. Data processed on the device does not leave it, which simplifies privacy reviews and consent. We still document what data is used and where, for your privacy policy and app store privacy labels.
Yes. Most projects add one or two AI features to an app that already exists, built in Swift, Kotlin, Flutter or React Native.
Yes. When a feature needs a cloud model, retrieval over company data or an AI agent, our AI team builds that part too and connects it to the app.
We build a test set of real examples, measure accuracy before release, and re-run it whenever the model, prompt or platform changes. Low-confidence results are shown with care or sent for review.
A feasibility check on real devices takes a few weeks. Adding a feature to an existing app usually takes a few weeks to a few months, depending on the model and how many devices you support.
We start with a short discovery call under NDA if you need one. Send us your idea, any designs or documents, and the systems the work must connect to. We then share a written scope, timeline and estimate before any build work starts.
Want AI features that are fast, private and work offline?
Book a call with our mobile AI team. We will review your idea, check it against your users' devices and suggest the best approach.