Ummaly.

A halal lifestyle app: scan a product to check it, find halal food and mosques nearby, prayer times, Qibla and more. Behind the scanner sits a verdict engine that combines product databases, curated rules, vector retrieval and an LLM, with versioned results so the same product always gets the same answer. I also trained my own halal classifier, measured it honestly against the live pipeline, and kept it in shadow mode until it is safe.

• Live on the App Store, plus a web app and a Next.js marketing site with a Three.js hero
• Fine-tuned GPT-2 Medium on a 7,634-example dataset built from public databases and hand-curated rules
• Model served by a FastAPI service in Docker; benchmarked against production and held in shadow mode rather than risk a false halal verdict
• User feedback and uncertain scans flow into an admin console that exports corrections for retraining