At Banque Misr, I built a LangGraph agent that turns business questions about loans and credit cards into SQL and Plotly charts. It retries failed queries and, when it can’t answer, suggests questions it can handle.
I added accuracy, correctness, and hallucination tests to CI/CD on every change, with an LLM judge evaluating each answer. I also tracked experiments in MLflow and traced system steps with OpenTelemetry to identify major cost and latency drivers.
At Banque Misr, I also created a zero-shot pipeline to classify customer complaints into a two-level taxonomy and quantified EGP 12M a year of revenue at risk. I fine-tuned an open-source LLM with LoRA and deployed it behind an agentic product search.
Earlier, I worked on a production voice assistant for Tawuniya, customer-service and ATM voice-assistant projects at Accord Business Group, and fraud, pricing, NLP, and computer-vision systems across previous roles.
