Mobile AI · On-Device Computer Vision
I build the whole pipeline: C++ image processing on the camera frame, an INT8 model that fits in kilobytes, and the Flutter app around it. One engineer, from raw sensor data to recognized text — no server in the loop.
ocr-demo.mp4 next to this file.
Why this is rare
From the camera frame to recognized text on one phone: a C++ NDK image pipeline, a quantized model running through LiteRT / TFLite, and the production Flutter app shipping it. People who can do all three are uncommon.
Mobile is a harsher environment than the server: a fixed power budget, no GPU farm, milliseconds that matter. I work in INT8, delegate to the NPU, and squeeze accuracy out of models small enough to ship. That discipline transfers straight to edge and embedded work.
I've built OCR for Arabic identity documents end to end — mixed right-to-left text, dates, ID numbers — the exact problem space of KYC, identity verification, and scanning products. Native Arabic, on top of the engineering.
Selected work
On-device OCR engine
A recognition engine for Arabic identity documents. Custom depthwise-separable CNN with a dual-head CTC decoder — one head for mixed Arabic text (names, addresses, RTL), one for pure numbers (dates, IDs) — quantized to INT8 and run fully on-device through LiteRT. The C++ pipeline handles colour conversion and preprocessing before a single frame ever reaches the model.
Mobile app · live
On-device face authentication built for business use. Camera capture, face processing, and full Flutter delivery on both Android and iOS — shipped through real store review and running on users' devices. The iOS build is distributed B2B, viewable on an iPhone or Mac.
Open-source · GitHub
A zero-dependency Flutter architecture core: a sealed Result<T> type for error handling and a Controller-as-State pattern, with 20+ tests. Built after the GetX maintenance gap left a hole for a small, dependable state layer.
Open-source package · pub.dev
Fast YUV→PNG conversion for camera and on-device vision pipelines — the kind of low-level frame plumbing every real-time camera feature needs and most projects re-write badly.
Toolkit