I'm a Data Scientist focused on building and evaluating AI systems. My work brings together machine learning, data pipelines, and software engineering, with a particular interest in natural language processing and retrieval-augmented generation (RAG).

I approach projects through clear hypotheses, controlled experiments, and inspectable results. I care about understanding why a model improves, which evidence supports an answer, and how someone else can reproduce the outcome.

• Applied machine learning: data preparation, experimental design, and model evaluation.
• NLP and language models: retrieval-augmented generation and research into model adaptation, grounding, and uncertainty.
• AI engineering: APIs, durable ingestion pipelines, retrieval infrastructure, and workflow orchestration.
• Reproducibility: versioned data, explicit configurations, retrieval traces, and documented experiment artifacts.

A local-first research engineering system, evolved from RAGForge, for computer science and AI experimentation. Its most developed area is an end-to-end RAG workflow that makes documents, retrieval evidence, and query history inspectable.

• Processes versioned documents through Bronze, Silver, and Gold data layers.
• Combines dense and sparse retrieval in Qdrant.
• Streams answers with ranked evidence and preserves query history.
• Includes retrieval tracing and orchestration benchmarks to support analysis of system behavior.

Technologies: Python, FastAPI, PostgreSQL, Qdrant, MinIO, Redis, Airflow, Celery, Next.js, Docker Compose.

Status: RAG capabilities are implemented; a broader framework for experiments across research domains is a planned direction.

An ongoing study of how fine-tuning and retrieval affect language-model answers in a medical question-answering dataset. The central research question is when a model should rely on adaptation, retrieve supporting evidence, or abstain from answering.

The planned evaluation compares four configurations: base Qwen, base Qwen with RAG, QLoRA-adapted Qwen, and QLoRA-adapted Qwen with RAG.

• Completed: data cleaning and group-based splitting designed to prevent data leakage.
• In progress: Qwen input formatting and token analysis.
• Planned: QLoRA training, retrieval-corpus construction, comparative evaluation, uncertainty analysis, and retrieval stress tests.

The study examines performance across common and rare topics, unsupported claims, and model behavior when evidence is irrelevant or contradictory.

I'm open to Data Science opportunities and collaborations in AI/ML engineering, NLP, applied AI, and model evaluation.

Explore my portfolio or contact me at elbaz.ouad1249@gmail.com.