Iyad Khalil

Fraud Detection System for Health Insurance Claims - Built an AI-first pipeline using GPT-4o (Azure OpenAI) to analyze 1,500+ prescriptions/day with p95 latency 900ms, enabling real-time compliance checks. - Implemented tool-calling JSON extraction (dosage, indications, contraindications, risks, etc.) from unstructured HTML posology, achieving 100% extraction accuracy across validation cases. - Integrated Blink Pharmacie API (6,000+ drugs) to enrich metadata and compute drug-drug interaction severity (low/med/high) aligned with pharmacist validation. - Designed anomaly detection & multi-drug risk workflows to support clinical decision-making; system validated against Moroccan drug standards and pharmacist review. - Tested on real-world fraudulent prescriptions with 100% detection accuracy, deployed as a FastAPI + Docker service on Azure with interoperable JSON/text reports and full auditability.

AI-Powered Revenue Forecasting Platform for Pharmacies - Built a hybrid LSTM + XGBoost forecaster for 35+ pharmacies over 6 years of daily transaction data, achieving 93.2% accuracy, SMAPE 8.7%, outperforming ARIMA, Prophet, SARIMA, AutoML baselines. - Served predictions via FastAPI + Docker on Azure (Linux VM) with p95 < 500ms latency and ultra-low infracost ($0.014/hr ≈ $0.002/request). - Optimized serving with batching + caching, reducing DB calls and cutting latency by ~40% under load. - Integrated a LangChain GPT assistant to translate forecasts into natural-language recommendations, validated by next-day actuals (high alignment). - Engineered 25+ domain-specific features (seasonality, volatility, patient behavior) to mitigate overfitting and improve generalization across new vs. older pharmacies. - Deployed with Dockerized reproducible Conda environments, model versioning, Postman monitoring; tracked performance for 4 months.