I build production ML systems that explain themselves.

Arabic Complaint Classification API

Three MARBERT-v2 classifiers — each one I fine-tuned myself — score sentiment, topic, and intent; a deterministic rule engine maps them to an action. The optional LLM layer only ever writes the explanation.

"The LLM explains, never decides."

Privacy-Preserving Family Resemblance

Given two parents' photos, find the children who most resemble the blended parental identity — entirely in embedding space. Faces become 512-d InsightFace vectors immediately; no photographs are ever stored.

Finalist — Social Development Innovation Hackathon · Team مُلائم

Legal RAG — Ask Saudi Law in Arabic

Embed the Arabic query, retrieve grounded chunks from Qdrant, and return a cited answer. If no LLM is configured, it degrades gracefully — returning retrieved context with an explicit LLM_NOT_CONFIGURED flag instead of failing.

Interpretable Cardiology Risk

Ten-year coronary risk on Framingham, run through three lenses: classical statistics, machine learning, and explainability (SHAP + Bayesian PyMC). The point isn't the model — it's which metric you trust on imbalanced clinical data.

Accuracy lies. Recall is what saves lives.