Jana Jouni

At InkyLab AI, I designed and built a multimodal video search system that retrieves events from long surveillance videos using natural-language queries. I combined video encoders with Qdrant indexing and added an InternVL3 re-ranking stage to reduce false matches. My video translation and voice cloning project started as an internship at Byblos Microsystems. I developed a system that transcribes and translates video subtitles, then reinserts bilingual subtitle files; I’m now extending it to generate dubbed audio through voice cloning. I built an automated C++ code grading system using Llama 3.1 70B, Ollama, and a Streamlit interface. In my tests, combining Few-Shot prompting with an Evaluator-Optimizer framework reached 94% grading accuracy. I also built an end-to-end fake news detection app with a GRU model and an automated MLOps lifecycle. The project uses MLflow and DVC for versioning, with a FastAPI backend, PostgreSQL storage, and GitHub Actions deployment pipelines.