Fayez Ali

Summary

What I did

• Built a persona-focused multilingual RAG chatbot for a Nobel Peace Prize holder. Used Azure (OpenAI Service, AI Search, Blob Storage, App Services), giving users conversational access to their writings, books, speeches, lectures, articles, and podcasts.
• Designed and implemented a credit risk scoring system for microfinance lending. Trained CatBoost, Random Forest, XGBoost, and LightGBM on the organization's historical borrower data to predict default probability and mapped outputs to an industry-standard 300-900 credit score. Integrated SHAP for model explainability to support loan underwriting and reduce dependence on manual risk assessment.
• Built an AI-powered meeting minutes generator using Azure OpenAI and Pyannote for speaker diarization, now adopted firm-wide and reducing documentation effort by ∼70%. Deployed the diarization service as a containerized inference endpoint on serverless GPU, isolating compute-intensive workloads from the main application on Azure App Service. Automated CI/CD with GitHub Actions.
• Developed a credit risk scoring model and a multilingual RAG chatbot.
• Deployed the complete credit risk scoring pipeline to the client's server.
• Developed a dashboard using FastAPI with endpoints for bulk scoring database records and retrieving stored results.
• Used prompting techniques to ensure LLM output accuracy by having the model check and correct its own format.

Results

• Reduced documentation effort by ∼70% through the firm-wide adoption of an AI-powered meeting minutes generator.
• Achieved about 82% accuracy for the credit risk scoring model.
• Optimized the credit risk model to improve AUC for better differentiation between good and bad borrowers.
• Tested the multilingual RAG chatbot with a maximum of 15-20 people.