CallMind

AI-powered Android app that turns phone call recordings into searchable, summarized conversation memory.

CallMind does not record calls. It watches a folder on your device (or imports files manually), transcribes audio, runs LLM analysis, and stores everything locally so you can browse, search, and export summaries.

Features

• Automatic discovery — foreground service monitors your recording directory via MediaStore and can auto-process new files
• Manual import — pick recordings from anywhere with the system file picker
• Transcription — cloud STT via Google Gemini multimodal (gemini-flash-lite-latest), or on-device Vosk (36 MB model download)
• Call analysis — summaries, key points, topics, sentiment, and checkable action items
• Semantic search — cloud (gemini-embedding-001) or on-device (MediaPipe Universal Sentence Encoder) vectors in Room, with cosine similarity and model-matched querying
• Text search — keyword search across transcripts and summaries
• Per-contact history — conversation threads grouped by contact
• Export — share call summaries as text files
• Flexible backends — Gemini or OpenAI-compatible LLM for analysis; cloud or local embeddings (switching providers triggers a full re-index)

How it works

New recording detected
        │
        ▼
┌───────────────────┐
│ TranscriptionWorker│  Gemini (cloud) or Vosk (on-device)
└─────────┬─────────┘
          ▼
┌───────────────────┐
│  AnalysisWorker   │  Gemini or OpenAI-compatible LLM
└─────────┬─────────┘
          ▼
┌───────────────────┐
│  EmbeddingWorker  │  Gemini or on-device embedder (user setting)
└─────────┬─────────┘
          ▼
   Room database + notifications

PipelineOrchestrator chains the three workers under a unique WorkManager name per call (pipeline_<callId>, REPLACE policy). Config errors (missing API key, model not downloaded) fail fast and are stored on the call row as processingError.

Recordings are matched to contacts using filename parsing and the Android call log. When matching fails, metadata is inferred from the filename alone.

Requirements

Android	API 29+ (Android 10+)
JDK	21 recommended (Gradle can auto-provision via Foojay)
Android SDK	API 36.1 (platforms;android-36.1) for current AGP 9.1
Device	Phone with existing call recordings (e.g. OEM dialer saves to Music/Recordings/Call Recordings/)
API keys	Gemini API key required for cloud transcription and cloud embeddings; optional second key/endpoint if using OpenAI-compatible analysis

Getting started

1. Clone and configure SDK

git clone <repo-url> CallMind
cd CallMind

Create local.properties in the project root (gitignored):

sdk.dir=/path/to/your/Android/sdk

2. Build

./gradlew assembleDebug

Install the APK from app/build/outputs/apk/debug/ on your device, or run from Android Studio.

3. First launch

1. Grant audio, call log, contacts, and notifications permissions.
2. Open Settings and set:
   • Recording directory — default: Music/Recordings/Call Recordings (adjust for your OEM)
   • Gemini API key — required for cloud STT and cloud embeddings (and Gemini-based analysis)
   • Analysis LLM — Gemini (default) or OpenAI-compatible provider
   • Speech-to-Text — cloud (default) or local Vosk (download the model for offline STT)
   • Search embeddings — cloud Gemini or on-device model (download required for local)
   • Auto-process — optionally process new recordings automatically
3. Use Scan on the home screen to index existing recordings, or enable auto-process and wait for new files.

Configuration reference

Setting	Purpose
Recording directory	Relative path under shared storage where your dialer saves WAV/other audio
Gemini API key	Cloud transcription, cloud embeddings, and Gemini-based analysis
Analysis LLM provider	Gemini or OpenAI Compatible (custom base URL, key, model)
Use local STT (Vosk)	On-device transcription; no Gemini needed for STT when enabled
Search embeddings	cloud (gemini-embedding-001) or local (MediaPipe on-device model)
Auto-process	Run the full pipeline when new recordings appear

Connection test buttons in Settings verify API keys and embedding providers before processing a backlog. Changing the embedding provider wipes the vector index and re-embeds analyzed calls so search only compares vectors from the same model.

Permissions

Permission	Why
READ_MEDIA_AUDIO	Read call recording files
READ_CALL_LOG	Match recordings to phone numbers and call times
READ_CONTACTS	Resolve contact names
POST_NOTIFICATIONS	Processing progress and completion alerts
FOREGROUND_SERVICE	Background monitoring and WorkManager jobs
INTERNET	Cloud STT, LLM, and embedding APIs

Development

Project structure

app/src/main/java/com/callmind/app/
├── CallMindApplication.kt       # Hilt + WorkManager
├── MainActivity.kt
├── di/                          # Hilt modules (Room, network)
├── data/
│   ├── local/db/                # Room v3 — calls, transcripts, analysis, embeddings
│   ├── local/preferences/       # DataStore settings
│   ├── local/                   # Vosk STT, embedding model download
│   ├── remote/                  # Gemini & OpenAI-compatible APIs, embedding providers
│   └── repository/              # CallRepository (single source of truth)
├── service/
│   ├── RecordingMonitorService  # ContentObserver + backlog scan
│   ├── PipelineOrchestrator     # WorkManager chain
│   └── worker/                  # Transcription, Analysis, Embedding
├── ui/                          # Compose — home, detail, search, contact, settings
└── util/                        # Recording parser, audio convert, semantic search, export

Common commands

Task	Command
Debug APK	./gradlew assembleDebug
Release APK	./gradlew assembleRelease
Unit tests	./gradlew testDebugUnitTest
Lint	./gradlew lint
Full check	./gradlew check

Instrumented tests (connectedDebugAndroidTest) need a connected device; CI typically runs unit tests and lint only.

Tech stack

• Kotlin · Jetpack Compose · Material 3
• MVVM · Hilt · Coroutines / Flow
• Room · DataStore · WorkManager
• Retrofit · OkHttp · Kotlin Serialization
• Vosk (local STT) · MediaPipe (local embeddings)

SDK note

This project uses AGP 9.1 with compileSdk 36.1. Install the matching platform in the SDK Manager:

sdkmanager "platforms;android-36.1"

Do not simplify the compileSdk block to compileSdk = 36 — the minor API level is required.

Codebase map (graphify)

A graphify knowledge graph can be generated locally for architecture exploration:

graphify extract .   # or full pipeline via /graphify in Claude Code
graphify query "How does CallRepository connect the pipeline?"

Outputs land in graphify-out/ (gitignored). See CLAUDE.md for agent-oriented build and architecture notes.

Privacy & data

• Recordings stay on your device; only audio/text you choose to process is sent to configured cloud APIs.
• API keys are stored in app-private DataStore preferences.
• Local embeddings run fully on-device after the model is downloaded.
• This is a personal / open-source project — not designed or audited for Play Store distribution.

Roadmap

Planned native upgrades (Apex whisper.cpp STT, EmbeddingGemma on-device search) are documented in tobeimplemented.md.

Near-term UI and quality work:

• Pull-to-refresh and richer error/retry UI on the home screen
• Optional dark/light theme toggle (currently follows system)
• Replace Vosk with Apex once whisper.cpp integration lands (see tobeimplemented.md)

License

No license file is included yet. Add one before public distribution if you plan to open-source the repo.