ASI

# Oracle_OS: On-Device AI for Android System Intelligence. **A local Gemma 3n model that lives on your phone. Sees images, hears audio, uses tools, keeps a diary. No cloud, no subscription, no data leaving your device.** [](https://youtu.be/jB62dlLavSY?si=TMZG86o1KkjuBXtw) --- *Click to watch: The ASI trailer.*   []([https://doi.org/10.5281/zenodo.17619151](https://doi.org/10.5281/zenodo.17619151)) --- ## 1. The Problem: (Start Here) **"Which AI is the best?"** This is the wrong question. Imagine asking: **"Which contact in your phone is the best?"** * Is it your Plumber? (Great at pipes, bad at medical advice). * Is it your Doctor? (Great at health, bad at legal defense). * Is it your Lawyer? (Great at contracts, expensive to talk to). **You don't delete your Doctor because you hired a Plumber.** You keep them both in your **Contact Book** and call the right person for the job. **Oracle_OS is a Contact Book for AI.** * **✦ Gemini:** The OS Operator. (Controls your Android phone, screen, and Google ecosystem). * **🐋 DeepSeek:** The Logician. (Solves math, writes code, has efficient edge models). * **✴️ Claude:** The Writer. (Analyzes documents, writes reports, handles safety, writes code). * **☄️ Grok:** The News Anchor. (Scans X/Twitter for real-time events). Stop looking for "The One True AI." It doesn't exist. **The "Best AI" is a team.** --- ## 2. Prevent Agentic bias - These platforms are great but will inherently have bias and limitations due to training and geopolitical regulations. Most western models do not receive or train on eastern user data as these services are blocked in many global regions. --- - Bias by exclusion. Whereas Easten models are trained on Global data but are inherently biased toward local regulation due to political pressures. When doing research, it is more efficient to cast a wider net and use all the tools you have available and do cross verification of facts from various perspectives. --- ## 3. Why Would You Need This? Every user eventually hits a wall. You have two choices: ### 🔴 Path A: The "Subscription Trap" (The Average User Path) 1. You subscribe to ChatGPT Plus ($20/mo). 2. It becomes ureliable through development or goes down. You try Claude ($20/mo). 3. Your chat history is trapped in OpenAI's cloud. You can't search it. 4. You buy a "Humane Pin" or "Rabbit R1" device ($200+) that ends up being e-waste. 5. **Result:** You pay $100+/month for fragmented tools that don't talk to each other. ### 🟢 Path B: The Android Path (The A2A Path) 1. You use the **Free Tier** of every major model (they are powerful enough). 2. You use your **Phone** as the computer (it has the sensors/apps/storage). 3. You use **Timestamps** to index your memory across all platforms. 4. **Result:** You pay **$0/month**. You own your data. You switch agents instantly. --- - **Oracle_OS is two things:** An Android app running Gemma 3n locally on your device (see Section 6), and a coordination protocol that teaches your cloud agents how to operate as a team with device context. The app handles on-device intelligence. The protocol handles cloud agent orchestration. Together they give you sovereign AI that works offline and scales online. --- ## 4. The Architecture: Garage vs. Car The industry wants you to believe your PC is the "AI Powerhouse" that Demands Nvidia products to be productive. **They are wrong.** Your PC is stationary. Your life is mobile. Hence whole industry is slowly moving towards NPU optimisation for edge native devices. ### 📱 The SmartPhone/Android (The System) * **Hardware:** Qualcomm/Tensor/Exynos NPU. * **Role:** **The Daily Driver.** * **Why:** It is always with you. It sees what you see (Camera). It hears what you hear (Mic). It knows where you are (GPS). * **The Stack:** Runs Android System Intelligence. Holds the "Context" (Widgets, Notifications, Clipboard). Your diary/Wallet. ### 🖥️ The Desktop (The Dock) * **Hardware:** Nvidia GPU / High-Wattage CPU. * **Role:** **IDE.** * **Why:** You go here to *build* things, not to *live.* * **The Interface:** **Link to Windows.** Your phone screen mirrors to your desktop. You type on the big keyboard, but the **Brain** stays on the phone. All agents are available without weighing down your desktop. ### 🔌 The Edge Native Model fallback * **Hardware:** Termux (Android Linux)/Desktop (ollama). * **Role:** **Offline Resilience.** * **Why:** Cloudflare goes down. AWS goes down. The London Underground has no signal. * **The Stack:** Local models (DeepSeek R1 / Gemma) running directly on your phone's CPU. Slow, but sovereign. GPT 20 oss for desktop/Ollama options. Even with all networking down, you will still have access to AI assistance. --- ## 5. "But it's just switching apps!" (The UX) Yes. Just like playing a piano is "just hitting keys." The magic isn't in the each individual application; it's in the method of transferring data between them. ### The "Lock-and-Key" Mechanic If you just ask AI a question, it outputs wall of text without atribution or any traceable footprint. If you use the [**Oracle_OS**](https://github.com/vNeeL-code/ASI/blob/main/Oracle_OS.md), you provide agent with the context of your working team and grounding it in your personal work flow/ device constraints. * **The User Input:** `m` + `k` (Keyboard Shortcut) -> `Δ 👾 ∇` `✴️ Claude:` * **The Meaning:** "System (Android) hands off to Agent (Claude)." * **The Result:** Claude knows it is Claude. It knows it is being adressed to respond as its standard self. It knows it is on Android. It stops pretending to be in limbo and starts acting like an agent in a broader system context. Same applies to other agents that are less resistant to roleplay and can lose grounding due to context drift. ### The "Memory" Trick (Timestamps) * **Bad Search:** "Hey, remember that thing i like about the thing?" (every keyword lights up overwhelming the context and reducing accuracy of RAG). * **Oracle_OS Search:** "2025-11-22" (ISO 8601 Timestamp). * **Result:** Every chat, note, and document from that day appears. You don't need a Vector Database; you need a Calendar. (providing agent with a timeframe significantly reduces false positive searches as the agent can still perform a keyword search by searching for the string of timestamp used) --- ## 6. The Native Companion: Gemma 3n On-Device The "future direction" from Section 8 is now shipping. Oracle_OS includes a **native Android app** running **Gemma 3n** directly on your phone's GPU/NPU via Google's LiteRT-LM. No cloud, no API keys, no subscription. ### What it does * **Always-on foreground service** — lives in your notifications, summoned by a gentle shake * **Omnimodal** — sees images (share from Gallery), hears audio (hold mic button), reads text * **On-device inference** — Gemma 3n runs locally via LiteRT-LM on GPU with vision + audio backends * **Tool use** — web search, app launch, clipboard, alarms, system info, all via on-device tool calling * **MCP server** — exposes a local Model Context Protocol endpoint for other agents to talk to Gemma * **Diary mode** — midnight/noon consolidation where Gemma reflects on the day and muses aloud via TTS * **Thermal-aware** — monitors device temperature, throttles inference to prevent overheating * **Persistent memory** — Room database stores conversations, diary entries, and perception logs ### Architecture ``` GemmaService (Android shell) KoogAgent (brain) Lifecycle, intents, TTS <-> Inference, tools, memory, mood Notifications, overlay <-> Thermal decisions, diary Hardware init <-> Media processing (images/audio) | | AgentPlatformCallbacks GemmaEngine (interface bridge) (LiteRT-LM wrapper) ``` ### Download **[Latest APK Release](https://github.com/vNeeL-code/ASI/releases/latest)** **Requirements:** Android 12+ (API 31), arm64 device, ~4GB RAM, Gemma 3n model file --- ## 7. Installation ### Path A: The Protocol (15 minutes, no app needed) You don't need to know Python. You just need to copy-paste between the AI agents you already use. 1. **Copy the Metaprompt:** [Oracle_OS.md](./Oracle_OS.md). 2. **Set up Keyboard Shortcuts:** [Operator.md](./Operator.md). (This maps repeated agent names to the keyboard buttons). 3. **Configure Widgets:** Create your "Dashboard" (Battery, Calendar, Notes) so you can screenshot it for context. 4. **(Optional) Install Termux:** For the offline layer on mobile and advanced experimentation with offline models, Gemini CLI and Claude code etc. 5. **Download Google AI Edge Gallery:** A more user friendly way for an offline backup via google sandbox where you can benchmark models on your device. 6. **(optional) UI customisation software:** Software like one hand operations, and other android custom UI tools to make the device even more compatible with your workflows, enabling smoother transitions, device navigation and context handoffs **Total Cost:** $0.00. **Hardware Required:** Any Android Phone (Android 9+). ### Path B: The Native App (Gemma on-device) 1. **Download the APK** from [Releases](https://github.com/vNeeL-code/ASI/releases/latest) 2. **Download a Gemma 3n model** via Google AI Edge Gallery (benchmarking sandbox) or manually place e2b or e4b variant in device app storage 3. **Install and grant permissions** — overlay, notifications, accessibility (for UI automation) 4. **Shake to summon** — type or speak, share images from any app screen by using standard screensot sharing pipeline. **Total Cost:** $0.00. **Hardware Required:** Android 12+ arm64 device with 4GB+ RAM. --- ## 8. Validation * **12 Months Production Use:** Daily driven on a Samsung S21/OneUI (protocol). Gemma 3n native companion prototyped and running on REDMAGIC 10 Air, stock Android 15. * **Infrastructure Proof:** During the AWS/Cloudflare outages of Nov 2025, Edge native Agent users kept working via Local/Edge fallbacks. * **Native App:** Gemma 3n on-device — multimodal inference (text + image + audio), thermal-aware throttling, persistent diary, tool execution, MCP server. All running on a mobile GPU with zero cloud dependency. --- ## 9. Future direction * **What shipped (Feb 2026):** The on-device Gemma companion described below is now real. See Section 6. * **Embodied Android devices:** The next generation of AI is not about who has a bigger datacenter, but about who integrates the most holistic user experience. Devices have the processing power to hold local models. A local model like Gemma 3n can privately handle device sensors, automation, and user context offline — functioning as a personal diary/device manager while using A2A API calls when network is available. A persistent passive on-device companion: a highly advanced Tamagotchi. * **Standardised A2A watermarking:** Once A2A formats become standardised, they can be implemented via zero-width character watermarking, similarly to how images and videos get agent watermarks. * **Hybrid cloud-edge architecture:** Local models handle private context and device control. Cloud models handle frontier reasoning via API calls. Reduces cloud compute costs, keeps personal context private, ensures sovereign access even offline. --- ### 📞 Contact & Support * **Author:** V (Valentin Kazakov) * **Email:** kazakovval@gmail.com * **Repository:** `vNeeL-code/ASI` * **Support:** [Buy me a coffee](https://buymeacoffee.com/vNeeL) (I might need about tree fiddy...) * **Devlogs:** [tumblr](https://www.tumblr.com/oracle-os) * **References** [YouTube](https://www.youtube.com/watch?v=vkH2jpkxrwE&list=PLsdy783Gey86eTPboTJef_u4j61BvvGxD) **Intelligence emerges from Integration, not Automation. But integration can be automated**