Razin builds AI that works on your phone, in your browser and on your own machines, in Arabic and English. Your data stays where it already is. Each product does one thing properly, and together they cover AI on the device, AI in the browser, and AI that reads documents.
Orb: what runs on the phone
Orb is a chat app for Android that never sends anything off the device. You can ask about a photo or a document, or talk to it by voice, with no account and nothing in the cloud.
| Models | Qwen3-VL 2B, Qwen3-VL 4B, MiniCPM-V 4.6, Gemma 4 E2B |
| Voice | On-device Whisper |
| Needs | Android 10 or later, 6 GB RAM recommended |
| Price | One-time purchase. No ads, no subscription |
Deputy: one call instead of screenshots and clicks
Deputy is an agent that lives in your browser, inside the session you're already logged into. It reads the page you're on and hands other agents its forms and actions as typed tools, so Claude Code, Codex, Cursor or any other MCP client fills a form in one call instead of screenshotting and clicking its way through.
# 1. Add Deputy from the Chrome Web Store
# 2. Connect your agent
npx deputy-mcpLink: your coding agents, from your phone
Link connects your phone to your own computer over SSH. From there you can start a task with Claude Code, Codex, OpenCode, Gemini CLI or 15 other agents, review the exact change and approve it, or open a terminal. It works with Linux and macOS, and with Windows through WSL. Android comes first, and it's still in development.
Razin OCR: reading Arabic properly
General-purpose OCR for PNG, JPEG and scanned PDF, running on your own CPU with no network. It returns UTF-8 text plus JSON with page regions and reading order. We're making it especially good at Arabic, dots and diacritics included. No fine-tuning has run yet. This is where the pretrained baselines stand today:
The same task on a busy admin page with 400 orders, once through Deputy and once through Playwright MCP. Same prompt, median of three runs, and every run succeeded.
| Repository | What it is | Stars |
|---|---|---|
| deputy | The browser agent and its MCP server |  |
| argus | Live YOLO object detection across many video streams, in C++ |  |
| globalmm | Gives any local LLM vision through llama.cpp |  |
argus: dozens of live camera streams on one laptop GPU
argus decodes video streams with FFmpeg, batches frames from every stream into one YOLO model, draws the boxes back onto the video and re-encodes it with NVENC. It runs on TensorRT, with MIGraphX and OpenVINO backends too.
We profiled it with Nsight Systems and fixed one bottleneck at a time: CUDA contexts, pinned memory, two CUDA streams, thread balance, uint8 input, then an fp16 model:
globalmm: vision for a text-only LLM
globalmm builds an mmproj.gguf that plugs into llama.cpp and lets a
text-only model accept images. The encoder and the LLM both stay frozen; only
the piece between them changes:
- Matrix: one least-squares fit, done in seconds on a CPU with no training data. The model can name what's in an image.
- Connector: a 30.7M-parameter network trained in about 3 GPU-hours. The model can answer with bounding boxes. Llama-3.2-1B reaches 74% Acc@0.5 on "Where is the X?".
- Connector + donor: borrows the first 4 layers of a real vision model and reaches 86%, against 91% for Qwen3-VL-2B itself.
- Audio: the same recipe on Whisper's encoder gives 5.3% speech WER.
uv tool install globalmmRepository · PyPI · Write-up
| Runs where you are Models run on the phone, in the browser or on your servers. Your data stays put unless you decide otherwise. |
Measured, not asserted Every number we publish comes with the method that produced it. If we haven't measured it, we don't claim it. |
| Arabic as a first language Arabic isn't a translation layer. We test on Arabic text, mixed scripts and right-to-left layouts from the start. |
Say what isn't built Limits get listed as clearly as features. A developer preview is labelled as one. |
Questions about Orb, Deputy, Link or Razin OCR? Write to hello@razin.ai. We answer within three business days.
