Principal AI / ML Engineer

I build AI that ships 🚢.

8+ years turning models into reliable production systems across the full ML stack.
Portfolio, research, and writing.

Karakoram, 2022

Hi, I'm Rafay 👋 — a Principal AI/ML Engineer based in and working out of Dubai & Islamabad.

Currently at VisionX, I'm the technical lead for AiCEO on Qi-Bot, an LLM-powered voice and chat assistant for Iraq's largest payments processor, Qi Card, serving millions of customers in Arabic, Kurdish, and English across multiple platforms. Before that, I worked with an amazing team at Veeve.io, where I led the ML Engineering team and and built the models behind "In-Cart Video Intelligence" — smart shopping carts running computer vision models (see an early prototype below). I've also helped ship edge inference SDKs in Rust at Lightly.ai, and spent two years as a research fellow at Mila, as part of the Fatima Fellowship, studying how data augmentation affects what models learn.

On the side, I'm building Skinsight.me with my friends — an agentic skincare recommender trained on 2M+ user signals across 60K products.

That's the work. If you wanna chat about any of it, shoot me a message 💌.

Featured Projects

Qi Card AI Assistant

Voice & chat LLM assistant for Iraq's largest payments network. Arabic, Kurdish and English, WhatsApp-first, real-time voice. Millions of active users.

LLMsVoice AIMultilingualSGLangAWS

In-Cart Video Intelligence

7-class action detection at 20 FPS on 1,500+ smart shopping carts across the US, Europe, and the Middle East. The demo above is an early prototype.

Computer VisionEdge AINvidia TritonKubernetesGCP

Hand-Object Segmentation

Personal project. Intel RealSense depth + RGB into a segmentation model for hands and held objects, running on-device at sub-10ms.

PyTorchRealSenseEdge AIONNX Runtime

Research Work

Mila

Research Fellow (part-time)Apr ’22 – Jun ’24

“More than more data? Comparing training with data augmentation vs IID data”

Designed controlled experiments comparing augmentation regimes against IID training to analyze effects on robustness, convergence, and learned representations. Developed custom augmentation methods and toy datasets to isolate how specific transformations affect model behavior. Built a modular Python experiment workbench for reproducible training. Key question: does more augmented data actually teach a model something different, or just the same thing louder?

Supervisor: Alex Hernandez GarciaCodebase

TUKL-NUST R&D Lab

Research Assistant / Vision LeadSep ’16 – Jun ’17

“Fish Biodiversity Estimation by Low-Cost Non-Destructive Video Based Sampling (FIBEVID)”

Led the computer vision team in collaboration with Pakistan Fisheries Department and University of Kaiserslautern to build a non-destructive fish biodiversity estimation tool for Pakistan’s National fish, the Mahasher. Project went on to receive funding from Hochschule RheinMain and DAAD (German Academic Exchange Service). Created and published the first-ever Mahasher fish dataset of 100,000 labeled images in its natural river habitat.

Supervisors: Dr. Faisal Shafiat, Dr. Imran MalikFIBEVID Project

TUKL-NUST R&D Lab

Research Assistant (part-time)Jun ’16 – Sep ’16

“Forgery detection and segmentation of handwritten text through Hyperspectral Image Analysis”

Summer intern collaborating with researchers at University of Kaiserslautern to detect forged bank cheques. Performed statistical analysis on hyperspectral image data to separate out inks of different varieties.

Supervisor: Dr. Imran Malik

From the blog

All posts