Starbuck Note Taker

Starbuck Note Taker is an Android note-taking app built with Kotlin + Jetpack Compose. It is designed to keep core note functionality available offline while storing note data locally in encrypted form.

What the current app does

• Create and edit notes with:
  • plain text and rich text formatting
  • checklist notes
  • optional event metadata (start/end, timezone, location)
• Attach images and files to notes.
• Generate note summaries on-device using MLC LLM with Llama 3.2 3B Instruct (q4f16_0 quantisation) via LlamaEngine; falls back to lightweight rule-based heuristics when the model is unavailable or the device has insufficient RAM.
  • Summarise — concise 1–3 line note preview
  • Rewrite — rewrites a note in a clean, professional style
  • Question — answers a free-form question using optional note context
• AI features require ≥ 4 GB total device RAM (DeviceCapabilityChecker); the rule-based fallback is used automatically on devices below this threshold.
• Add URL link previews (metadata fetch requires internet once; cached preview data remains local afterward).
• Protect notes with PIN-based encryption, with optional biometric unlock support.
• Schedule event reminders and full-screen alarm flows.
• Accept shared content from other apps via Android share intents (SEND, SEND_MULTIPLE).
• Support in-app audio transcription using Android SpeechRecognizer (availability/quality depends on device speech services and permissions).

Security and storage model

• Notes are saved in encrypted local storage (notes.enc) using AES/GCM with a key derived from the user PIN.
• Attachments are stored locally and referenced from note records.
• Locked notes remain hidden until unlocked with the configured PIN/biometric flow.

Offline behavior

The app is offline-first for note creation, editing, storage, reminders, and summarization.

Network may still be used for:

• fetching new link-preview metadata,
• speech recognition depending on the device’s speech engine configuration,
• downloading Llama 3.2 3B model weights (~2 GB) on first use (stored locally under filesDir/models/ afterward).

Project structure

app/
  src/main/java/...                Android app source
  src/main/assets/                 Bundled ML assets + deployment docs
  src/main/assets/scripts/         Offline model training + verification scripts
scripts/                           Repository helper scripts

Build and test

During Codex development, run Gradle with --console=plain.

./gradlew test --console=plain
./gradlew assembleDebug --console=plain

AI / ML setup

MLC model library (required before building)

The compiled Llama 3.2 3B model library must be placed in app/src/main/jniLibs/arm64-v8a/ before building. There is no prebuilt Android .so for this model in the public MLC release artifacts, so you must compile it locally:

# Prerequisites:
# - Python 3.10+
# - Android NDK r27+
# - pip install mlc-llm
# - export ANDROID_NDK=~/Android/Sdk/ndk/<version>
# - export TVM_NDK_CC=$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android24-clang

# 1. Download the quantised weights
git clone https://huggingface.co/mlc-ai/Llama-3.2-3B-Instruct-q4f16_0-MLC

# 2. Compile the model library for Android arm64
mlc_llm compile \
  ./Llama-3.2-3B-Instruct-q4f16_0-MLC \
  --target android \
  --device android:arm64-v8a \
  -o Llama-3.2-3B-Instruct-q4f16_0-MLC-android.so

# 3. Copy into the project (note the required "lib" prefix)
cp Llama-3.2-3B-Instruct-q4f16_0-MLC-android.so \
   app/src/main/jniLibs/arm64-v8a/libLlama-3.2-3B-Instruct-q4f16_0-MLC.so

The .so is the compiled model library (~MB range). Model weights (~2 GB) are downloaded automatically at runtime from HuggingFace on first use.

Legacy TFLite assets (optional)

The files note_classifier.tflite, tokenizer_vocabulary_v2.txt, category_mapping.json, and deployment_metadata.json in app/src/main/assets/ are legacy artifacts from the previous TFLite classification path. They are no longer used at runtime (the rule-based fallback fills this role now) but are retained for reference. To regenerate them:

python3 app/src/main/assets/scripts/complete_pipeline.py