feat: Update version and add Hive models for profile management

- Bump version to 1.7.10+11 in pubspec.yaml
- Introduce Hive models for ProfileCard, ActivityLog, ChatMessage, AgentMemory, and DossierEntry to manage user profiles and chat history.
- Implement serialization and deserialization methods for the new models.
- Create ChatBubble widget for displaying chat messages with support for images and thinking content.
- Add ChatHistoryModal for managing chat history and profiles.
- Implement ChatInputBar for sending messages with image attachment support.
- Develop DebugPanel to show LLM stats and performance metrics.

Author: princepspolycap

README.md

@@ -44,11 +44,14 @@ A micro-app that runs **100% on your device** - no cloud, no data sharing. Just
 
 ## Features
 
-### 🤖 Local AI (Tiered Intelligence)
-- **Featherweight (270M)** - Ultra-fast agentic tools & actions.
-- **Middleweight (1B - 1.5B)** - High-quality conversational chat (Gemma 3, DeepSeek R1).
-- **Heavyweight (3.8B)** - Advanced reasoning (Phi-4 Mini).
-- **Multimodal (Vision)** - Screenshot analysis for interest grading.
+### 🤖 On-Device AI (5 Curated Models)
+| Model | Context | Capabilities |
+|-------|---------|--------------|
+| **Gemma 3n E2B** 📷 | 32K | Vision + Audio + Tools |
+| **DeepSeek R1** 🧠 | 32K | Reasoning + Thinking + Tools |
+| **Phi-4 Mini** 🔬 | 128K | Best context window + Tools |
+| **Gemma 3 1B** 💎 | 32K | Recommended chat quality |
+| **Qwen 2.5 1.5B** 🌍 | 32K | Multilingual + Tools |
 
 ### 🔒 100% Private
 - All AI runs on-device via **LiteRT (MediaPipe)**.

docs/MODELS.md

@@ -1,96 +1,72 @@
 # Wingman LLM Models
 
-## Architecture: Multi-Model Intelligence
+## Architecture: Curated Model Selection
 
-Wingman supports a tiered model strategy to balance speed, intelligence, and privacy.
+Wingman uses a curated set of 5 high-quality models focused on:
+- **Good context windows** (32K-128K tokens)
+- **Multimodal capabilities** (vision, audio)
+- **Function calling / tools** support
 
-### The Intelligence Tiers
+### Model Summary
 
-| Tier | Size (Params) | Best For | Recommended Model |
-|------|---------------|----------|-------------------|
-| **Featherweight** | 270M - 0.5B | Speed & Tools | **FunctionGemma 270M** ⭐ |
-| **Middleweight** | 1B - 1.5B | Chat & Reasoning | **Gemma 3 1B** 💎 |
-| **Heavyweight** | 3B - 4B | Advanced Logic | **Phi-4 Mini** 🔬 |
+| Model | Params | Size | RAM | Context | Capabilities |
+|-------|--------|------|-----|---------|--------------|
+| **Gemma 3n E2B** 📷 | 2B eff | 3.6GB | 4GB | 32K | Text + Vision + Audio + Tools |
+| **DeepSeek R1** 🧠 | 1.5B | 1.2GB | 4GB | 32K | Text + Thinking + Tools |
+| **Phi-4 Mini** 🔬 | 3.8B | 2.8GB | 6GB | 128K | Text + Tools (Best context!) |
+| **Gemma 3 1B** 💎 | 1B | 689MB | 4GB | 32K | Text (Recommended for chat) |
+| **Qwen 2.5 1.5B** 🌍 | 1.5B | 1.1GB | 4GB | 32K | Text + Tools (Multilingual) |
 
 ---
 
-## Model Roster (Verified Official List)
+## Model Details
 
-All models are official LiteRT builds from the **`litert-community`** organization on HuggingFace.
+### Multimodal
+*   **Gemma 3n E2B** 📷🎤: Full multimodal - text, image, video, audio input. Best for vision tasks. Requires Gemma license acceptance on HuggingFace.
 
-| Model | Class | Size | RAM | Context | Category |
-|-------|-------|------|-----|---------|----------|
-| **DeepSeek R1** | 1.5B | 1.2GB | 4GB | 32K | Reasoning/Agent |
-| **Phi-4 Mini** | 3.8B | 2.8GB | 6GB | 128K | Logic/Reasoning |
-| **Gemma 3 1B** 💎 | 1B | 689MB | 4GB | 32K | General Chat |
-| **Gemma 3 270M** | 270M | 276MB | 2GB | 32K | Fast Chat |
-| **Qwen 2.5 1.5B** | 1.5B | 1.1GB | 4GB | 32K | Multilingual |
-| **Qwen 2.5 0.5B** | 0.5B | 400MB | 2GB | 32K | Fast Actions |
-| **Gemma 2 2B IT** | 2B | 1.6GB | 4GB | 8K | High Quality Text |
-| **FunctionGemma** ⭐ | 270M | 150MB | 2GB | 8K | Specialized Agent |
-| **TinyLlama 1.1B** | 1.1B | 650MB | 2GB | 2K | Legacy Support |
-| **SmolLM 135M** | 135M | 270MB | 1GB | 2K | Ultra-Lightweight |
+### Agent & Reasoning  
+*   **DeepSeek R1 1.5B** 🧠: Thinking mode + function calling. Best for complex reasoning and tool chaining.
+*   **Phi-4 Mini** 🔬: Advanced logic with 128K context - the largest context window available. Best for long conversations.
+*   **Qwen 2.5 1.5B** 🌍: Strong multilingual support + tools. Best for non-English users.
 
----
-
-## Model Tiers
-
-### Tier 1: Agent & Tools (The "Brain")
-*   **DeepSeek R1 1.5B**: Best for complex reasoning and tool chaining.
-*   **Phi-4 Mini**: Advanced logic for deeper strategy analysis.
-*   **FunctionGemma 270M**: The core model for simple on-device actions.
-*   **Qwen 2.5 (1.5B/0.5B)**: Exceptional for multilingual tool requests.
-
-### Tier 2: General Chat (The "Vibe")
-*   **Gemma 3 1B** 💎: **Recommended.** Best balance of wit, speed, and persona.
-*   **Gemma 2 2B IT**: High-precision text for nuanced advice.
-*   **Gemma 3 270M**: Whisper-fast chat for quick check-ins.
-
-### Tier 3: Support Models (The "Safety Net")
-*   **TinyLlama 1.1B**: Reliable chat for older hardware.
-*   **SmolLM 135M**: Runs on almost anything; perfect for emergency advice.
+### General Chat
+*   **Gemma 3 1B** 💎: **Recommended.** Best balance of quality, speed, and size. Requires Gemma license.
 
 ---
 
-## Technical Setup
-
-### RAM Recommendations
-- **Old Devices (2GB)**: Use FunctionGemma 270M.
-- **Standard (4GB)**: Gemma 3 1B or DeepSeek R1.
-- **Pro Devices (6GB+)**: Phi-4 Mini or Gemma 3n E4B.
+## RAM Recommendations
 
-### Context Management
-Wingman uses a dynamic sliding-window context. Older messages are summarized once they exceed the model's `contextWindowK` (listed above), ensuring you nunca pierdas el hilo de la conversacion.
+- **Standard (4GB)**: Gemma 3 1B (chat) or DeepSeek R1 (reasoning)
+- **Pro Devices (6GB+)**: Phi-4 Mini (128K context) or Gemma 3n E2B (multimodal)
 
 ---
 
 ## Official Sources & Repositories
 
-Wingman uses strictly verified `.task` models to ensure compatibility with MediaPipe/LiteRT.
+All models are verified `.task` or `.litertlm` formats for MediaPipe/LiteRT compatibility.
 
-### Primary Source: LiteRT Community
-Most models are sourced from the [LiteRT Community HuggingFace Organization](https://huggingface.co/litert-community).
-- **DeepSeek R1 1.5B**: [Link](https://huggingface.co/litert-community/DeepSeek-R1-Distill-Qwen-1.5B)
-- **Phi-4 Mini**: [Link](https://huggingface.co/litert-community/Phi-4-mini-instruct)
-- **Gemma 3 Family**: [Link](https://huggingface.co/litert-community/Gemma3-1B-IT)
-- **Qwen 2.5 Family**: [Link](https://huggingface.co/litert-community/Qwen2.5-1.5B-Instruct)
+### LiteRT Community (Public - No login required)
+- **DeepSeek R1 1.5B**: [litert-community/DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/litert-community/DeepSeek-R1-Distill-Qwen-1.5B)
+- **Phi-4 Mini**: [litert-community/Phi-4-mini-instruct](https://huggingface.co/litert-community/Phi-4-mini-instruct)
+- **Qwen 2.5 1.5B**: [litert-community/Qwen2.5-1.5B-Instruct](https://huggingface.co/litert-community/Qwen2.5-1.5B-Instruct)
 
-### Specialized Agent Models
-- **FunctionGemma 270M**: Sourced from [sasha-denisov/function-gemma-270M-it](https://huggingface.co/sasha-denisov/function-gemma-270M-it). This model is specifically fine-tuned for on-device tool calling and low-latency interaction.
+### Google (Gated - Requires Gemma license acceptance)
+- **Gemma 3n E2B**: [google/gemma-3n-E2B-it-litert-lm](https://huggingface.co/google/gemma-3n-E2B-it-litert-lm)
+- **Gemma 3 1B**: [litert-community/Gemma3-1B-IT](https://huggingface.co/litert-community/Gemma3-1B-IT)
 
 ---
 
-## Technical Specifications (Summary)
+## Technical Specifications
 
 | Metric | Minimum | Recommended |
 |--------|---------|-------------|
-| **Device RAM** | 2GB | 4GB - 6GB |
+| **Device RAM** | 4GB | 6GB+ |
 | **Storage** | 1GB | 5GB+ (for multiple models) |
 | **OS** | iOS 16.0+ / Android 11+ | Latest |
 
 ---
 
 ## Credits
 - **Google AI**: For the Gemma architecture and MediaPipe SDK.
-- **DeepSeek / Microsoft / Alibaba**: For the weights used in distilled LiteRT models.
-- **Sasha Denisov**: For the FunctionGemma fine-tune optimized for mobile agents.
+- **DeepSeek / Microsoft / Alibaba**: For model weights used in LiteRT conversions.
+- **LiteRT Community**: For maintaining verified on-device model conversions.

docs/ROADMAP.md

@@ -82,11 +82,18 @@ Local AI dating assistant with on-device LLM, voice input, and CRM-style profile
 ## Phase 2 - Intelligence & Language Learning
 
 ### Screenshot Analysis (Vision Model)
-- [x] Gemma 3n E2B multimodal support (already integrated)
+- [x] Gemma 3n E2B multimodal support (vision + audio)
 - [ ] Auto-detect language in screenshots
 - [ ] Extract conversation context
 - [ ] Interest level scoring
 
+### Curated Model Selection (5 models)
+- [x] Gemma 3n E2B - Multimodal (32K context)
+- [x] DeepSeek R1 1.5B - Reasoning + Thinking (32K context)
+- [x] Phi-4 Mini - Best context window (128K context)
+- [x] Gemma 3 1B - Recommended chat (32K context)
+- [x] Qwen 2.5 1.5B - Multilingual + Tools (32K context)
+
 ### Agentic Features (Tool Calling)
 - [ ] Parse function call responses
 - [ ] Auto-copy responses to clipboard
@@ -133,13 +140,17 @@ Local AI dating assistant with on-device LLM, voice input, and CRM-style profile
 ├─────────────────────────────────────────────┤
 │  LlmService (Dart)                          │
 │       ↓                                     │
-│  flutter_gemma (Native Bridge)              │
+│  flutter_gemma 0.11.14 (Native Bridge)      │
 │       ↓                                     │
 │  MediaPipe GenAI (C++)                      │
 │       ↓                                     │
 │  TensorFlowLite (Metal GPU on iOS)          │
 │       ↓                                     │
-│  Gemma 3n E2B (~1.5GB from HuggingFace)     │
-│  • Text chat • Vision • Audio • Tools       │
+│  5 Curated Models from HuggingFace:         │
+│  • Gemma 3n E2B (Vision/Audio/Tools)        │
+│  • DeepSeek R1 (Reasoning/Thinking)         │
+│  • Phi-4 Mini (128K context)                │
+│  • Gemma 3 1B (Recommended chat)            │
+│  • Qwen 2.5 1.5B (Multilingual)             │
 └─────────────────────────────────────────────┘
 ```

docs/wip/flutter_gemma_integration_fixes.md

@@ -394,11 +394,9 @@ Need to verify each `AvailableModel` has the correct `ModelType` assignment:
 |-------|-------------------|-------------------|
 | gemma3nE2b | gemmaIt | gemmaIt ✓ |
 | deepSeekR1 | deepSeek | deepSeek ✓ |
-| functionGemma270m | functionGemma | functionGemma ✓ |
-| qwen32bQ4 | qwen | qwen ✓ |
-| phi4Mini | general | general (verify) |
-| llama32_1b | llama | llama ✓ |
-| hammerQ4 | hammer | hammer ✓ |
+| phi4Mini | general | general ✓ |
+| gemma3_1bIt | gemmaIt | gemmaIt ✓ |
+| qwen25_15b | qwen | qwen ✓ |
 
 ### Bug #6: fileType Not Being Set
 
@@ -503,13 +501,11 @@ The library distinguishes between:
   - Pay special attention to Phi-4 (uses `general`)
 
 - [ ] **6. Test each model type after fixes**
-  - Test Gemma 3n E2B (gemmaIt)
-  - Test DeepSeek R1 (deepSeek + isThinking)
-  - Test Qwen 3.2B (qwen)
-  - Test Function Gemma (functionGemma)
-  - Test Llama 3.2 1B (llama)
-  - Test Hammer (hammer)
-  - Test Phi-4 (general)
+  - Test Gemma 3n E2B (gemmaIt) - Multimodal
+  - Test DeepSeek R1 (deepSeek + isThinking) - Reasoning
+  - Test Phi-4 Mini (general) - Best context
+  - Test Gemma 3 1B (gemmaIt) - Recommended chat
+  - Test Qwen 2.5 1.5B (qwen) - Multilingual
 
 ---
 
@@ -538,7 +534,11 @@ fileType: _currentConfig?.fileType ?? ModelFileType.task,
 
 ### Testing
 
-After implementation, test each model type (gemma, deepseek, qwen, llama, hammer, phi) to verify they work.
+After implementation, test each model type:
+- gemmaIt: Gemma 3n E2B, Gemma 3 1B
+- deepSeek: DeepSeek R1 (+ isThinking)
+- qwen: Qwen 2.5 1.5B
+- general: Phi-4 Mini
 
 ### Don't
 

ios/Podfile.lock

@@ -1,6 +1,4 @@
 PODS:
-  - adaptive_platform_ui (0.1.0):
-    - Flutter
   - background_downloader (0.0.1):
     - Flutter
   - CwlCatchException (2.2.1):
@@ -87,7 +85,6 @@ PODS:
   - TensorFlowLiteSwift/Privacy (0.0.1-nightly.20250619)
 
 DEPENDENCIES:
-  - adaptive_platform_ui (from `.symlinks/plugins/adaptive_platform_ui/ios`)
   - background_downloader (from `.symlinks/plugins/background_downloader/ios`)
   - file_picker (from `.symlinks/plugins/file_picker/ios`)
   - Flutter (from `Flutter`)
@@ -116,8 +113,6 @@ SPEC REPOS:
     - TensorFlowLiteSwift
 
 EXTERNAL SOURCES:
-  adaptive_platform_ui:
-    :path: ".symlinks/plugins/adaptive_platform_ui/ios"
   background_downloader:
     :path: ".symlinks/plugins/background_downloader/ios"
   file_picker:
@@ -144,7 +139,6 @@ EXTERNAL SOURCES:
     :path: ".symlinks/plugins/speech_to_text/darwin"
 
 SPEC CHECKSUMS:
-  adaptive_platform_ui: f54f620666d1797636f9ef06f49137ef2cc13611
   background_downloader: 50e91d979067b82081aba359d7d916b3ba5fadad
   CwlCatchException: 7acc161b299a6de7f0a46a6ed741eae2c8b4d75a
   CwlCatchExceptionSupport: 54ccab8d8c78907b57f99717fb19d4cc3bce02dc

ios/version.txt

@@ -1 +1 @@
-10
+11

lib/main.dart

@@ -10,10 +10,11 @@ import 'theme/app_theme.dart';
 
 import 'screens/chat_screen.dart';
 import 'screens/settings_screen.dart';
-import 'models/profile_card.dart';
+import 'models/hive_models.dart';
 import 'models/user_psychograph.dart';
 import 'services/storage_service.dart';
 import 'services/logger_service.dart';
+import 'services/notification_service.dart';
 
 void main() async {
   WidgetsFlutterBinding.ensureInitialized();
@@ -30,6 +31,8 @@ void main() async {
   Hive.registerAdapter(ActivityLogAdapter());
   Hive.registerAdapter(UserPsychographAdapter());
   Hive.registerAdapter(PatternEntryAdapter());
+  Hive.registerAdapter(AgentMemoryAdapter());
+  Hive.registerAdapter(DossierEntryAdapter());
 
   // Initialize storage service (opens boxes)
   final storageService = StorageService();
@@ -38,6 +41,9 @@ void main() async {
   // Initialize FlutterGemma plugin
   await FlutterGemma.initialize();
 
+  // Initialize notification service for reminders
+  await NotificationService.instance.initialize();
+
   log.i('Wingman app initialized');
 
   runApp(