Rafael Avraham

I built an open-source retrieval-augmented generation (RAG) chatbot with a Telegram interface for accurate, context-aware answers from custom data sources. I designed the end-to-end architecture using PostgreSQL with pgvector for semantic search, multi-agent orchestration for retrieval and response generation, OpenAI/Groq integrations for flexible inference, Docker and Google Cloud Run for deployment, Alembic migrations, and Langfuse for observability, tracing, and prompt versioning. The result is a production-ready GenAI application that keeps operational costs low while preserving scalability and maintainability.

I built a computer vision proof-of-concept for a penalty-training device that detects the instant before a player takes a penalty and triggers a randomized target signal in the goal. I designed the ML workflow around video analysis, notebooks, model artifacts, and Python inference utilities, turning sports-performance data into an interactive training aid for decision-making and reaction practice.

I handled a computer vision project to track and count people in video streams using YOLOv8 and the Supervision library. The project demonstrates end-to-end object detection, tracking, and counting workflows in a Jupyter notebook environment, showing how modern detection models can transform raw video into structured operational insights.