Hady Ibrahim

I’ve been lucky to work across both product engineering and applied ML, and I enjoy roles where I can go from a fuzzy problem to something people can actually use. Early on, I worked at Apple and at Shopify in software-focused roles where I learned the fundamentals of shipping: clean interfaces, thoughtful debugging, and building for the long-term instead of the demo. More recently at Shopify, I moved deeper into ML systems work at scale. I delivered a brand-recognition model that raised feature F1 from 12% to 73% and improved the overall model score by 39%, with gains statistically validated via bootstrapped 95% confidence intervals. To make supervision trustworthy, I built a stratified synthetic-data pipeline with LLM annotators and arbitrators, distilled features from a GPT teacher into a Qwen2.5-VL-7B student model and published versioned synthetic train/test sets. I deployed the model across real-time and streaming services to predict brand for 2+ billion products, modernized data products with historical and “latest” dbt prediction tables, and introduced an LLM-based judge to generate diagnostics and seed GRPO training/evaluation datasets. In research, I’ve been working on ultrasound-based microrobot detection for future microrobotic surgery. Ultrasound is rotation-heavy and noisy by default, so robustness and efficiency aren’t optional. I’m benchmarking steerable/equivariant CNN approaches that encode rotational structure into the model so it doesn’t need to relearn the same feature at every orientation, and that work is moving toward a manuscript. I’m also pushing the same “real constraints” mindset through audio in my current Capstone. We’re building an on-device selective hearing (AI hearing aid) prototype that can lock onto a chosen speaker in real time without sending audio to the cloud. Our pipeline pairs a noise-robust speaker enrollment module (a compact voice “fingerprint” from a short sample) with a target speech extraction model, and we’re adapting TF-GridNet from a non-causal research design into a causal, streaming variant so it can run continuously for real-time inference. Outside the lab, I’ve grown with McMaster’s Formula Electric team from team member to leading a 15-person Software & Embedded Systems group responsible for the car’s software stack. Along the way, I switched onto the Electrical team as well, where I gained hands-on PCB design, bring-up, and testing experience. Formula taught me how to build systems that survive integration, deadlines, and reality—where good documentation, clear interfaces, and rigorous testing make the difference between something that works once and something the whole team can trust.