Jandal AI

Play Store release in progress. Pre-release builds are available from GitHub Releases.

Jandal AI is a local-first Android assistant. It combines on-device chat, long-term memory, deterministic Android actions, voice input/output, and native tool calling while keeping user data on the device by default.

How it works

Jandal is built around a Brain-Memory-Action model:

• Brain - Gemma-4 E-2B / E-4B runs locally through Google AI Edge LiteRT / LiteRT-LM. A lightweight QuickIntentRouter handles deterministic fast paths for common device actions.
• Memory - local Room storage plus sqlite-vec / EmbeddingGemma support semantic recall, conversation history, core memories, and episodic summaries.
• Action - native Kotlin skills execute Android actions such as alarms, timers, lists, notes, weather, media controls, messages, email, calendar, navigation, Wikipedia, and unit/currency conversion.

Current launch status

Launch readiness is tracked in GitHub rather than duplicated in this README:

• #1014 - Play Store Launch Readiness & QA is the parent launch epic and current launch tracking issue.
• The launch:blocking label and Jandal Launch Backlog project views are the live launch blocker queue and filtered work dashboard.
• docs/PLAN-launch-slice.md is a historical launch-slice snapshot, not the current plan.
• #1255 - Launch Plan: ordered implementation sequence and release gates is the now-completed model-readiness preflight tracker, not the active launch plan.

The current launch gates are:

1. Launch scope and backlog sanity.
2. Test/evidence foundation.
3. Core app, accessibility, and first-run reliability.
4. Permissions and Android capability repair.
5. Voice and wake-word launch-risk validation.
6. Final release QA and store readiness.

Feature areas

Jandal currently includes:

• local chat with markdown rendering, multi-conversation management, streaming generation, and model/runtime status;
• local memory with core memories, episodic summaries, semantic search, and memory management screens;
• deterministic Android skills for lists, notes, alarms, timers, date arithmetic, media controls, unit/currency conversion, weather, messaging/email flows, navigation, and Wikipedia;
• drawer-accessible tools such as Lists, Alarms, Notes, Meal plans, and Settings;
• deterministic meal-planning sessions with plan approval, recipe persistence, favourites, replacement/regeneration, and ingredient export to lists;
• push-to-talk voice input, optional spoken chat replies, per-message speaker playback, configurable TTS voices, and wake-word/default-assistant infrastructure;
• local model management and model availability states for required, optional, preparing, gated, and unavailable model paths.

Future roadmap work includes Dreaming Engine background consolidation, Wasm skill extensibility, broader device optimisation, and richer multimodal/vision workflows. These are intentionally separated from the launch plan unless explicitly promoted.

Tech stack

Area	Technology
Language	Kotlin
UI	Jetpack Compose, Material 3 Dynamic Color, Glance widgets
Dependency injection	Hilt
Persistence	Room, DataStore
Background work	WorkManager
Local inference	Google AI Edge LiteRT / LiteRT-LM
Chat models	Gemma-4 E-2B / E-4B LiteRT-LM packages
Embeddings / RAG	EmbeddingGemma-300M, sqlite-vec
Quick actions	Kotlin QuickIntentRouter with deterministic routing and slot filling
Tool calling	LiteRT-LM native @Tool annotations and app-owned native skills
STT	Android native STT, Vosk, Sherpa-ONNX Zipformer / SenseVoice / Whisper tiny.en / Paraformer
TTS	Sherpa-ONNX Piper/VITS voice packs, Android TTS fallback, Kokoro research path
Wake word	Sherpa-ONNX verification and ONNX Runtime wake-word pipeline
Auth for gated downloads	Hugging Face OAuth via AppAuth
Extensibility research	Chicory WebAssembly runtime

Models and downloads

Model files are not committed to this repository. Some model downloads are gated and require a Hugging Face account plus acceptance of the upstream model licence/terms before the app can download or use them.

See models/README.md for the current model file reference, approximate sizes, ADB setup notes, and device-specific guidance.

Key launch-relevant examples:

Model / asset	Approx. size	Notes
Gemma-4 E-2B LiteRT-LM	~2.4 GB	Required launch-compatible chat model tier
Gemma-4 E-4B LiteRT-LM	~3.4 GB	Optional flagship-tier chat model
EmbeddingGemma 300M	varies by file	Required for local embedding/RAG paths where enabled
Sherpa STT models	~72-220 MB each	Downloaded per selected STT engine
Sherpa Piper/VITS voice packs	~64-116 MB each	Downloaded per selected voice; Semaine launch decision tracked in #1258

Privacy model

Jandal is designed as a local-first assistant:

• chat, memories, notes, lists, and model execution are local by default;
• production builds should not include telemetry;
• external network calls are limited to explicit web-backed skills or model/voice downloads where configured;
• Hugging Face sign-in is used only for gated model access.

Documentation

• models/README.md - model file setup, sizes, and ADB paths.
• docs/LEGAL_AND_ATTRIBUTION.md - launch attribution and licence checklist.
• docs/SPECIFICATION.md - broader product and architecture specification.
• docs/ROADMAP.md - broader roadmap notes.
• docs/testing/ - automated test harness, evidence schema, and QA documentation.

Licence and attribution

The source code is licensed under the Apache License 2.0. See LICENSE.

Third-party notices are maintained in NOTICE. Launch readiness for third-party libraries, voice assets, and gated model downloads is tracked in docs/LEGAL_AND_ATTRIBUTION.md.