Starbuck Note Taker

Starbuck Note Taker is an offline-first Android note-taking app built with Kotlin and Jetpack Compose. Core notes, checklists, reminders, attachments, and protected-note workflows remain local to the device.

Features

• Create rich-text, checklist, and event notes with attachments.
• Protect selected notes and archived collections with PIN-based encryption and optional biometric unlock.
• Schedule reminders and full-screen alarm flows.
• Accept shared text and files through Android share intents.
• Transcribe audio through Android SpeechRecognizer.
• Generate summaries, rewrite notes, and answer questions with one on-device Qwen model.
• Perform assistant web research on-device when a question needs current public information. See on-device web research.

AI architecture

Qwen3 0.6B is the only semantic and generative model used by the app. The model runs through Google LiteRT-LM:

• SUMMARISE creates grounded, category-aware note previews.
• REWRITE corrects and restructures a note without changing protected facts.
• QUESTION keeps every note context discrete, uses /note for explicit note evidence, and synthesises bounded public web evidence.

The canonical system prompts live in config/AI_AGENT_PROMPTS.txt. Its [AI_SUMMARISER], [AI_CHATBOT], and [AI_REFORMATTING] sections are copied into the APK during preBuild and loaded by AiAgentPrompts.

The pinned mixed-int4 model, qwen3_0_6b_mixed_int4.litertlm, is downloaded from litert-community/Qwen3-0.6B on first use and verified for expected size and SHA-256 checksum. The download is approximately 475 MB and is stored under the app's private filesDir/models/ directory. ARM64 devices try the GPU backend before CPU; x86_64 emulator builds use CPU. Devices need at least 4 GB of total RAM to load the model.

There is no TensorFlow Lite note classifier and no MLC/TVM compiler step in the application build. The Android dependency graph uses com.google.ai.edge.litertlm:litertlm-android.

When Qwen is unavailable, summary surfaces may show a bounded plain-text preview, rewriting returns the original text, and questions report that the model is unavailable. These results are fallback UI behavior, not output from a second AI model.

Privacy and network behavior

Notes and attachments are stored locally. Notes are saved in encrypted local storage (notes.enc) using AES/GCM with a key derived from the user PIN.

Network access may be used for:

• the one-time Qwen model download;
• public assistant web research;
• link-preview metadata;
• speech recognition, depending on the device speech service.

Private note content remains on-device. The app performs public page discovery and extraction, then passes bounded evidence to Qwen. Related-note retrieval excludes locked notes unless the user has unlocked them for the current process.

Project structure

app/src/main/java/...       Android application and Qwen/LiteRT-LM integration
app/src/main/assets/...     APK asset documentation
config/AI_AGENT_PROMPTS.txt Canonical Qwen system prompts
docs/...                    Architecture and feature documentation

Build and test

JDK 17, Android SDK 34, build tools 34.0.0, and NDK 26.1.10909125 are required.

./gradlew test --no-daemon --console=plain
./gradlew assembleDebug --no-daemon --console=plain

The manual deploy.yml workflow runs the same standard Gradle APK build. It does not install Python AI packages or compile model-native artifacts.