ahmed elsayed

On WhatsApp Twin, I diagnosed a DPO run that reached 100% preference accuracy by step 36 as a data shortcut. I rebuilt the dataset with hard negatives from my own replies and raised held-out accuracy from 68.7% to 71.3%, with a significant McNemar test result.

I built a pure-Python data pipeline that turns a WhatsApp export into multi-turn training samples, with PII redaction, a session-level split to prevent leakage, and 11 unit tests. I also designed a paired evaluation across unseen conversations using an LLM judge in both A/B orders.

For LLM Engineer's Handbook, I built a 68-question retrieval evaluation set and studied a production architecture with separate feature, training, and inference pipelines. I deployed a fine-tuned 3B DPO model to an AWS SageMaker endpoint behind a FastAPI microservice, and worked on the project's GitHub Actions CI/CD and Opik prompt monitoring.

My Mini-RAG course project was a FastAPI service for document upload, semantic indexing, and retrieval-augmented question answering. I moved its storage from MongoDB to PostgreSQL with PGVector. I'm looking for freelance ML auditing and evaluation projects.