Peer-reviewed publications

Green AI & Sustainable Machine Learning Research

Published as Muhammad Ahmad Zia — researcher and lecturer at the University of Lahore. My work investigates how machine learning systems can deliver equal or better results at significantly lower compute, energy, and cost, with applications across sustainable computing, smart cities, and commercial AI workloads. Many projects have open-source code available on GitHub.

🎓 Google Scholar 📑 Scopus 🔬 ResearchGate 🆔 ORCID 💻 GitHub

Green AI

Can One Setting You've Never Touched Cut AI Training Energy in Half? We Tested It on 100,000 Patient Drug Reviews

Journal of Computing & Biomedical Informatics (JCBI), Vol. 11, Issue 02, 2026 — Scopus Q3

A new peer-reviewed study takes a parameter that ships with every gradient boosting library and almost nobody changes, drops it, and measures what happens. On two biomedical drug-review datasets, reducing CatBoost's feature bin count to 64 cut training time and CO2 by roughly 48% while a formal equivalence test confirmed accuracy was statistically indistinguishable from the default. The same one-line change saved 25 to 60 percent in LightGBM and XGBoost.

Read summary → Full paper ↗ Code ↗

Does Lowering Precision Actually Speed Up AI Inference? We Ran 6,000 Benchmarks to Find Out

In Press

A new in-press study benchmarks gradient-boosted tree inference across four free-tier cloud platforms, three engines, and five precision levels using 6,000 measured runs. The surprising result: shrinking floating-point precision barely helps, naive INT8 quantization can quietly wreck accuracy, and engine choice — not hardware — explains nearly 80% of the speed difference.

Read summary →

Health-Aware AI

Would a Travel App Send an Asthma Patient to a High-Pollution City? We Tested 81 Destinations to Find Out

Journal of Computing & Biomedical Informatics (JCBI), Vol. 11, Issue 02, 2026 — Scopus Q3

A new peer-reviewed study rebuilds a travel recommender so that altitude, UV, and air quality are part of the ranking instead of an afterthought, then measures the carbon cost of the model class chosen to run it. The finding: recommendations for travelers with chronic conditions move away from clinically contraindicated destinations at no accuracy cost, and the industry-default 200-tree Random Forest burns roughly 22 times the training emissions of XGBoost while scoring slightly worse.

Read summary → Full paper ↗ Code ↗

Is 'Suitable for Sensitive Skin' Actually Safe? We Tested 1,263 Skincare Products to Find Out

Future Cities Conference 2026 (FCC'26), Healthy Cities Track

A new in-press study adds a transparent, regulator-grounded safety layer to skincare recommenders, tested on 1,263 real products and a live Pakistani marketplace. The finding: today's suitability labels barely protect shoppers past the first page of results, but a re-ranker built entirely from public regulator lists cuts allergen exposure by up to 99 percent at almost no cost to recommendation quality, and every flag it raises traces back to a named public source.

Read summary →

Research That Feeds Real Client Work

Every finding above informs how I build for clients — and every client deployment feeds evidence back into the research.

Book Free Discovery Call →