Example apps

DVAI-Bridge ships 18 runnable example apps. Every (SDK × backend) pair is covered. Each example is a complete project — its own manifest (package.json, Package.swift, build.gradle.kts, pubspec.yaml, .csproj) and a smoke.sh that boots a wiring-only test you can run on CI.

Source: examples/ in the monorepo. Grouped by SDK below.

Prerequisites (all examples)

# One-time, from the repo root:
git clone https://github.com/dvai-global/dvai-bridge.git
cd dvai-bridge
pnpm install --ignore-scripts

Native examples also need the native SDK published locally:

# Android: publish the workspace AAR to ~/.m2/
pwsh scripts/android-publish-local.ps1    # Windows
bash scripts/android-publish-local.sh     # Mac / Linux

# iOS: build the llama.cpp xcframeworks once (on macOS)
bash scripts/mac-side-prepare-xcframework.sh

Web / JavaScript

web-react — React + Vite + Transformers.js (browser, MSW)

A React chat UI. Talks to LangChain ChatOpenAI pointed at the local MSW endpoint. Default backend is Transformers.js. Swap to WebLLM if you want.

pnpm --filter web-react dev          # Vite dev server with HMR
# → open http://localhost:5173, click the prompt, watch tokens stream.

Verify: tokens appear in the response panel within ~5s after the model downloads (~500 MB on first run, cached afterwards).

Adapt: change transformersModelId in src/App.tsx to swap models. Swap @langchain/openai for the official openai package if you prefer.

web-vanilla-cdn — single HTML page, no bundler (browser, MSW)

One HTML page. A <script type="module"> tag pulls @dvai-bridge/vanilla from a CDN, calls initialize(), then fetch() against the local MSW endpoint.

( cd examples/web-vanilla-cdn && python -m http.server 8000 )
# → open http://localhost:8000
bash examples/web-vanilla-cdn/smoke.sh   # static smoke (no model load)

Adapt: edit index.html directly. No build step.

node-langchain — Node + Transformers.js + LangChain

dvai-bridge running in plain Node. LangChain's ChatOpenAI.stream() calls http://127.0.0.1:38883/v1.

pnpm --filter node-langchain start
# First run downloads the model into the HuggingFace cache.

Verify: streaming completion tokens print to the terminal.

Adapt: change modelId in index.js. Rewire the LangChain prompt or chain however you like.

node-llama-cpp — Node + native llama.cpp + LangChain

dvai-bridge with backend: "native". Loads a GGUF checkpoint via node-llama-cpp's NAPI bindings.

pnpm --filter node-llama-cpp start
# First run downloads ~800 MB GGUF.
bash examples/node-llama-cpp/smoke.sh    # gated by SMOKE_MODEL_URL / SMOKE_MODEL_SHA256

Verify: token rate roughly matches your CPU / GPU's tok/s ceiling.

Adapt: any GGUF that llama.cpp loads will work. Swap the URL and SHA-256 in index.js.

iOS native

Every iOS example opens with open Package.swift. Pick iPhone 16 as the run destination.

Example	Backend	Model	Open
ios-llama/	llama.cpp	Llama-3.2-1B-Instruct Q4_K_M (~800 MB)	cd examples/ios-llama && open Package.swift
ios-foundation/	Apple Foundation Models	(Apple-managed)	cd examples/ios-foundation && open Package.swift
ios-coreml/	CoreML	finnvoorhees/coreml-Llama-3.2-1B-Instruct-4bit	cd examples/ios-coreml && open Package.swift
ios-mlx/	MLX	mlx-community/Llama-3.2-3B-Instruct-4bit	cd examples/ios-mlx && open Package.swift

Build all four headlessly (on macOS):

bash scripts/mac-side-build-examples.sh build

Smoke any one in CI:

bash examples/ios-llama/smoke.sh
# On Mac: runs xcodebuild test asserting /v1/chat/completions returns non-empty.
# Off Mac: prints skip and exits 0.

Verify (after running on the simulator): tap Load + Ask. Watch the status label go through Download → Load → Bind. Tokens stream into the scroll view.

Adapt: edit Sources/<App>/<App>.swift. Change modelUrl, modelSha256, or swap the OpenAI SDK — the URL is bound.baseUrl.

Android native

Every Android example builds with Gradle. They consume the workspace AAR from mavenLocal().

Example	Backend	Model
android-llama/	llama.cpp	Llama-3.2-1B-Instruct Q4_K_M (~770 MB)
android-mediapipe/	MediaPipe LLM	Gemma-2-2B-IT .task (~1.3 GB)
android-litert/	LiteRT	litert-community/Llama-3.2-1B-Instruct.tflite (~1.0 GB)

# Publish the AAR (one-time, after SDK source changes):
pwsh scripts/android-publish-local.ps1

# Build + smoke one example:
cd examples/android-llama && bash smoke.sh
# Runs JVM tests; connectedAndroidTest auto-skips when no device.

Run interactively:

cd examples/android-llama
./gradlew :app:installDebug && adb shell am start -n co.deepvoiceai.examples.androidllama/.MainActivity

Verify: tap Load + Ask. Watch the Compose UI step through the states. Tokens render in the text view.

Adapt: edit MainActivity.kt. Change modelUrl / modelSha256. Swap aallam/openai-kotlin for any other OpenAI client pointed at bound.baseUrl.

React Native

react-native-app — backend selector, all six native backends

One screen. Dropdown to pick a backend. Supply modelPath, call DVAIBridge.start(...), stream a chat completion through the official openai npm SDK.

Run on Android (any host):

pnpm --filter react-native-app android

Run on iOS (Mac only):

pnpm install --ignore-scripts
cd examples/react-native-app/ios && bundle install && bundle exec pod install
cd .. && pnpm --filter react-native-app ios

Smoke:

bash examples/react-native-app/smoke.sh
# Typecheck + Metro bundle. RUN_ANDROID_BUILD=1 opts into the Gradle build.

Verify: the in-app log panel prints licenseStatus.kind and the bound baseUrl. The response area streams tokens.

Adapt: edit App.tsx. The backend dropdown values map to native enum cases. The model-path text input feeds StartOptions.

Flutter

flutter-app — backend dropdown, streaming via dart:io HttpClient

One StreamBuilder<DVAIBridgeState> drives the UI. The user picks a backend. DVAIBridge.instance.start(...) boots the embedded server.

cd examples/flutter-app
flutter pub get
flutter run                # picks the connected device / simulator
bash smoke.sh              # pub get + pigeon regen + analyze + test

Verify: the dropdown disables platform-mismatched backends in place. Pick mlx on Android — you get a DVAIBridgeError(backendUnavailable).

Adapt: edit lib/main.dart. Swap the backend defaults or the prompt template.

.NET

Example	Type	Backend	Host
dotnet-maui/	MAUI single-project	Auto / Llama / Foundation / CoreML / MLX / MediaPipe / LiteRT	Android any; iOS / Catalyst Mac-only
dotnet-desktop-llama/	Console + Avalonia	llama.cpp via DVAIBridge.Desktop	Win / Mac / Linux
dotnet-desktop-onnx/	Console + Avalonia	ONNX Runtime GenAI via DVAIBridge.OnnxRuntime	any desktop
dotnet-desktop-mlnet/	Console only	ML.NET OnnxScoringEstimator	any desktop

Build + smoke any one:

cd examples/dotnet-desktop-llama && bash smoke.sh
# Wiring-only verification; no model required.

Run interactively:

cd examples/dotnet-desktop-llama
dotnet build -c Release
dotnet run -c Release             # UI mode (Avalonia)
DVAI_HEADLESS=1 DVAI_MODEL_PATH=/path/to/model.gguf dotnet run -c Release

Verify (UI mode): the Avalonia window shows the bound base URL. Click Send. The streaming completion fills the text panel.

Adapt: every .csproj <ProjectReference> resolves into packages/dvai-bridge-dotnet/. Change the model path or backend selector in Program.cs.

Capacitor

capacitor-mobile — hybrid web bundle + native llama.cpp

A single HTML page in www/ calls DVAIBridge.start({ backend: "llama", modelPath }) from the webview. The chat input streams a completion via fetch() SSE.

# Web bundle:
pnpm --filter capacitor-mobile build

# One-time per platform:
pnpm --filter capacitor-mobile exec cap add android       # any host
pnpm --filter capacitor-mobile exec cap add ios           # Mac only

# Sync after every change to www/ or to a workspace plugin:
pnpm --filter capacitor-mobile exec cap sync android
# iOS (on macOS):
pnpm --filter capacitor-mobile exec cap sync ios
( cd examples/capacitor-mobile/ios/App && pod install )

# Run:
pnpm --filter capacitor-mobile exec cap run android
pnpm --filter capacitor-mobile exec cap run ios     # macOS only

Smoke:

bash examples/capacitor-mobile/smoke.sh
# Builds the web bundle and runs `cap doctor`. No device required.

Verify: the webview loads. The chat textarea accepts a prompt. The response area streams tokens.

Adapt: www/index.html is a plain HTML page — edit it directly. The scripts/build-www.mjs esbuild runner picks up the changes.

CI smoke matrix

Every example has a smoke.sh. It:

1. Skips cleanly when its host-OS / hardware requirement isn't met.
2. Runs a wiring-only test on supported hosts — typecheck, bundle, dotnet build, Gradle JVM tests, xcodebuild test against a wiring assertion.
3. Optionally exercises a real model when a SMOKE_MODEL_URL / SMOKE_MODEL_PATH env var is provided.

# Run every smoke that the current host supports:
for ex in examples/*/smoke.sh; do
  echo "=== $ex ==="
  bash "$ex"
done

The full per-(SDK × backend) matrix with host requirements lives in examples/MATRIX.md.

See also

• Getting started — the human-readable quickstart.
• License setup — what to add before shipping any of these to production.
• Per-SDK guide: iOS, Android, React Native, Flutter, .NET.