Jandal AI

Play Store release — coming soon. Pre-release builds on GitHub Releases.

How It Works

The app operates on a Brain–Memory–Action triad using a three-tier Resident Agent Architecture:

• The Brain: Gemma-4 E-4B/E-2B runs resident on GPU via LiteRT. A lightweight QuickIntentRouter (regex + MiniLM fallback) handles instant device actions and slot-filling fast paths. Complex queries go straight to Gemma-4 for full reasoning with native tool calling.
• The Memory: A local RAG (Retrieval-Augmented Generation) system using sqlite-vec and EmbeddingGemma-300M. The assistant remembers personal facts, preferences, and conversation history across sessions with zero data leaving the device. Episodic distillation consolidates each conversation into long-term memories.
• The Action: A modular skill framework. Tier 2 native Kotlin actions execute instantly (torch, timer, DND, bluetooth, lists, date arithmetic). Tier 3 complex skills (weather, calendar, memory recall, Wikipedia) are handled by the resident Gemma-4 model via LiteRT-LM's native @Tool annotations with SDK constrained decoding and rich inline result cards. Tool instructions are injected per turn only when a request looks tool-oriented, which keeps normal chat prompts slim as the skill set grows. Community-extensible WebAssembly skills run sandboxed via Chicory for safe extensibility.

Tech Stack

Component	Technology
Language	Kotlin
UI	Jetpack Compose, Material 3 Dynamic Color
Inference	Google AI Edge (LiteRT + LiteRT-LM)
Reasoning	Gemma-4 E-4B / E-2B (INT4 quantized, GPU resident)
Quick Actions	QuickIntentRouter (Kotlin regex, zero memory)
Complex Tool Calling	LiteRT-LM native @Tool annotations + constrained decoding
Embeddings	EmbeddingGemma-300M (768-dim)
Vector Search	sqlite-vec (NDK)
Wasm Runtime	Chicory (pure JVM)
DI	Hilt
Persistence	Room
Min SDK	API 35 (Android 15)
Speech-to-Text	Vosk (default), Android SpeechRecognizer (optional)
Text-to-Speech	Android TTS (default), Sherpa Piper/Kokoro (optional)

Features

Delivered

• 🧠 On-device reasoning — Gemma-4 E-4B running on GPU via LiteRT, no internet required
• 💾 Persistent memory — RAG-powered recall across conversations using sqlite-vec semantic search
• 🔒 100% private — No cloud APIs, no telemetry, all data stays on device
• 💬 Full markdown rendering — headings, bold, italic, inline code, code blocks, tables, links, lists
• 🎯 Smart chat titles — auto-generated from conversation content
• 🗂️ Multi-conversation — create, delete, rename, and search conversations by title
• 📁 Conversation management — archive, pin, drag-to-reorder conversations; swipe-to-archive (left) or delete (right) with confirmation; multi-select with bulk archive/restore/delete; auto-delete archived conversations after configurable retention period (default 7 days)
• 🧠 Core memories — add and manage permanent facts the assistant always recalls
• 📂 Memory Management screen — view, add and delete core memories and episodic memory browser
• ⚙️ Model selection — switch between E-2B and E-4B in Settings
• 🩴 Persona modes — choose Full Jandal, Half a Jandal (default), or Boring AI Mode in Model Management
• 🎬 Fun loading screens — 13 themed animated narratives
• 🖼️ Context window management — structured prompt assembly with KV cache management and recursive summarisation
• 📊 Runtime info — shows active model, backend (GPU/NPU/CPU), and device tier in chat
• ⚡ Quick Actions tab — instant device commands (torch, timer, DND, bluetooth) via zero-overhead Kotlin pattern matcher
• 💭 Episodic memory distillation — Gemma-4 summarises each conversation into long-term memories
• 🔧 Native skills — alarms, timers, SMS, email, torch, calendar events, weather (GPS + city), navigation, media playback, Wikipedia
• 📆 Native date arithmetic — deterministic get_date_diff handling for "days until/since" queries
• 🔍 search_memory — semantic search across core + episodic memories on demand
• 🌦️ Weather follow-up resolution — colloquial weather phrasing plus indirect references like "there" and "the capital of New Zealand"
• 🧩 Quick intent slot filling — supported fast-path intents can pause for missing required parameters instead of failing silently
• 🪪 Rich tool presentations — weather cards, list cards, confirmation chips, expandable previews, and surfaced fallback links
• 🔎 Tool call debugging — expand any tool call chip to see request/result, tap to copy
• 🧭 Nav drawer — Lists, Alarms, and Meal plans accessible from Chat, Actions, and all main screens via hamburger menu
• 📋 Lists — create and manage named lists via chat ("add milk to shopping list") or the Lists UI; full CRUD with active/completed sections
• 🗒️ Notes — create, edit, and manage free-form notes via chat or the Notes UI; smart auto-titles, pin/unpin, archive/unarchive, search, sort, drag-to-reorder, multi-select bulk archive/delete/restore/pin/share, and archived-only view
• 🗓️ Scheduled Alarms — date-specific alarms scheduled via Jandal appear in the Alarms screen for review and cancellation
• 📟 Side panel — slide-out drawer accessible from Chat and Settings shows active alarms and timers with live countdown; cancel any from the panel
• 🎵 Media controls — pause, stop, skip, and previous track via Jandal ("skip song", "pause music")
• 🎙️ Podcast playback — open a podcast app and start/resume playback ("play my podcasts", "resume podcast")
• ⏱️ Timer management — list active timers and cancel individual ones ("cancel my 10 minute timer")
• 🗑️ Alarm multiselect delete — select multiple alarms in the Alarms screen and delete them at once
• 📝 Bulk list add — add multiple items to a list in one request ("save all ingredients to shopping list")
• 🗣️ Offline voice controls — push-to-talk Quick Actions plus spoken responses managed from Settings → Voice
• 🔊 Streaming spoken chat replies — chat TTS begins playback before generation completes; preprocessing layer handles URL colon preservation, speech rate clamping, and abbreviation-aware sentence splitting (KNOWN_ABBREV + INITIALS_REGEX) so "Dr.", "Mr.", "e.g." don't break sentences; Sherpa voice quality evaluated and tuned on device
• 🔊 Per-message speaker button — VolumeUp icon on every assistant bubble; tap to play or stop that message's TTS independently of voice mode
• ⚙️ Expanded TTS settings — pitch slider (Sherpa only, 0.5–2.0×), auto-speak chat replies toggle (decoupled from Quick Actions via autoSpeakEnabled field), max spoken sentences dropdown (0 = unlimited, 2, 3, 5); all grouped in a "Chat voice behaviour" section in Settings
• 🛑 Verbal stop command — saying "stop", "stop speaking", "cancel", "be quiet", "shut up", or "silence" during TTS playback cancels speech and stops mic re-arm
• 🗣️ VCTK multi-speaker selection — choose from 109 VCTK voices (gender filter, speaker ID, accent label) in Settings → Voice; sid mapping sourced directly from the Piper model config
• 🗣️ Semaine multi-speaker selection — 4 distinct voices (Prudence, Spike, Obadiah, Poppy) selectable in Settings → Voice (#818, PR #818)
• 📐 Deterministic unit conversion — length, mass, volume, temperature, speed with alias normalisation and spoken-STT variants (#676, PR #816)
• 💱 Deterministic currency conversion — ISO code resolution via Frankfurter/ECB rates with same-currency short-circuit and clear error for unsupported currencies (#831, PR #848)
• 🗣️ TTS pronoun normalisation — converts first-person pronouns (my/I → your/you) in spoken summaries so the assistant speaks in third person (#828, PR #830)
• 🔊 Voice fallthrough preservation — Actions→Chat fallthrough preserves the user's voice-speak expectation so replies are spoken even after cross-screen navigation (#832, PR #833)
• ⏱️ Slot-fill retry on no-speech — system retries the slot-fill prompt instead of failing; cancel phrases abort the flow; start-listening audio cue confirms mic activity (#790/#791, PR #825)
• 🔒 Blank response guard — retries without RAG before showing fallback when LiteRT produces 0 tokens; keeps chat awake during load and generation (#839/#841, PRs #840/#842)
• 🎙️ Homescreen Glance widget — quick actions and voice from the launcher via GlanceAppWidget; VoiceCommandActivity and WidgetTextInputActivity with task isolation (#617, PR #847)
• 🔧 Audio quality fixes — AudioTrack tail cutoff prevention via hardware-latency silence padding; expectedSlotPromptSpeech normalisation to match TTS output; SID=0 clamp for single-speaker voices; aye pronunciation correction (#837/#828/#810, PRs #838/#836/#811)
• 🍽️ Deterministic meal planner — app-owned meal-planning sessions with bounded JSON generation, draft-plan approval, progressive recipe reveal, visible x of y progress, interruption-safe resume, quantity sanity validation, cuisine preferences, Kiwi wording normalization, batch day replace/regenerate, and quick-action/chat handoff (#859/#869/#931/#932/#971, PRs #864/#875/this PR)
• 📚 Meal plans browser — drawer-accessible Recent plans and Favourites tabs with recipe search, canonical favourite toggles, recipe re-add to Lists, and ingredient export into existing user lists (#933, PR #934)
• 💬 Multi-turn dialog — expanded confirmation, digression, and slot-filling coverage across more intents (#708, PR #712)
• ⏰ Alarms CRUD UI — create, edit, and toggle alarms directly from the Alarms screen; full CRUD via nav drawer (#479, PR #484)
• 🗒️ Lists management upgrades — rename, pin, sort, edit items, favorites, and due dates (#662)

Coming Soon

• 🗒️ Lists — hierarchical items — nested sub-items within lists (#928)
• 🌙 Dreaming Engine — overnight WorkManager consolidation (Light Sleep → REM → Deep Sleep) (Phase 4)
• ⚡ Semantic cache — instant responses for repeated knowledge queries (Phase 4)
• 🪪 Self-healing identity — structured user profile, LLM-managed via Dreaming cycle (Phase 4)
• 🧩 Wasm skill store — community-extensible plugins with sandboxed execution (Phase 5)
• 🏠 Home Assistant / Google Home — smart home control (Phase 5)
• 📱 8GB device optimisation — dynamic weight loading/unloading, E2B fallback (Phase 6)
• 🎙️ "Hey Jandal" wake word — always-on local detection → instant action routing (Phase 3F)

Roadmap

Work is organised by launch priority against the Google Play Store release. The sequencing below follows the slices in docs/PLAN-launch-slice.md, which is the living plan derived from the full backlog review. Labels: 🔴 launch:blocking must ship before publish · 🟡 launch:post follows shortly after · ⚪ launch:deferred is consciously parked.

Counts as of the last backlog review: 24 blocking · 45 post · 17 deferred. Tracked under Epic #1014 — Play Store Launch Readiness & QA.

🔴 Launch Blocking — ordered by delivery slice

Slice 1 · Foundation & stability (do first — unblocks everything)

Issue	Size	Summary
#915 · #916	L · S	Toolchain upgrade — AGP 9 / Gradle 9 / Kotlin 2.3.x / Hilt (touches every module)
#428	M	Memory profiling — peak RAM & concurrent model usage (feeds #430/#432)
#692	M	Fix inference stalls in Boring AI Mode
#937 · #957	M · S	Memory + intent-routing correctness bugs

Slice 2 · Heavy hitters — memory-safe model lifecycle (highest risk)

Issue	Size	Summary
#430	XL	Dynamic model loading state machine — never hold Gemma-4 + EmbeddingGemma at once
#432	L	Compatibility-tier model swap — auto E-2B + smaller KV cache on 8GB devices

Slice 3 · Navigation & visual quality (store-listing readiness)

Issue	Size	Summary
#747	M	Back-button & blank-screen navigation bug
#751	L	Navigation refactor — surface Lists / People / Clock / Settings
#226	L	Jandal visual identity — palette, loading states, 🩴 treatment
#961	M	In-chat model settings controls

Slice 4 · Finish in-flight capabilities

Issue	Size	Summary
#996	S	Wire Sherpa-ONNX as wake-word dual-threshold verify window
#885 · #886	M · M	Messaging — reply via RemoteInput; send to named group chats
#261	M	Skill discoverability
#928	L	Hierarchical list items
#713	L	Vision foundation — single-image Q&A + image-in-chat
#756 · #824	M · M	Voice — Piper voice-training research; Phase 3F on-device QA gate

Slice 5 · Release gate (run last)

Issue	Size	Summary
#427	XL	Comprehensive verification — full feature matrix on physical S23 Ultra
#868	S	Documentation & licence/attribution review
#441	M	Publish to Play Store — account, signing, listing, policy compliance

🟡 Post-Launch — fast-follow after publish

• Memory & data — cosine-distance vec tables (#647), anaphoric "remember that" (#958), low-confidence search filtering (#959), Artifact entity (#235)
• Skills — WebSearch (#407), calendar events (#942), podcast quick actions (#587), map/location (#258), Plex API (#594) & YouTube Music (#596)
• Voice — STT Kiwi/Māori normalisation epic (#935), translator skill (#659), TTS HW accel (#852), Kokoro quality (#854), wake-word FP tuning (#986)
• Chat UX & vision — thinking-message refactor (#964), Ephemeral Vision pipeline (#287), multimodal audio input (#943)
• Optimisation & runtime — Matryoshka 256-dim embeddings (#429), battery deferral (#431), tok/s benchmark (#803), grounded-numeric reliability (#968)
• Test harness — UIAutomator coverage (#548) plus harness hardening (#554, #560, #562, #563)

⚪ Deferred — parked behind larger phases

• Phase 4 — Dreaming Engine (#705) incl. graph-DB memory research (#419)
• Phase 5 — Wasm Runtime + Skill Store (#706) incl. MCP integration research (#944)
• Phase 6 — Device Optimisation (#707) incl. S21 generation failure (#684)
• Model experiments (#704) — Qwen 3.5 4B/0.8B (#691, #699), llama.cpp backend (#702)
• Alternative STT — Parakeet CTC (#700) & whisper.cpp (#703) — gated on the Sherpa-default decision (#1008)
• Fun / content skills — joke (#819), storytelling (#820), learn-something-new (#949)

Full sequencing, dependencies, and critical path: docs/PLAN-launch-slice.md.