Amine Chebil

• Developed an AI multi-agent email assistant based on LangGraph and LangChain, designed to help hotels analyze client emails, generate accurate responses, reduce response time and improve customer service. • Designed and developed two agentic workflows: - A sequential pipeline for email analysis implementing profanity detection, automatic summarization, and multi-label classification across 45+ hotel categories with dynamic flow control. - A multi-agent workflow for response generation integrating a supervisor agent orchestrating two specialized sub-agents: RAG agent and web search agent. • Integrated Qdrant as a vector database and Hugging Face (all-MiniLM-L6-v2) embeddings for accessing hotel documents and contextual search. • Integrated web tools with Tavily Search to answer general knowledge questions. • Improved generation of multilingual emails with configurable parameters (tone, length, theme, language) and automatic structured output via Pydantic. • Implemented the Subagents pattern to replace the Handoffs pattern, resolving the sequential delegation issue and restoring parallel agent execution. • Exposed the system using FastAPI and integration with Outlook for native email selection. • Monitored and evaluated agents via LangSmith, including interaction tracing, output analysis, Prompt Engineering optimization, and agent performance improvements. -Technologies: Python, LangGraph, LangChain, LangSmith, RAG, FastAPI, Qdrant Cloud, Docker, Groq, Hugging Face Embeddings, Tavily Search, Pydantic

• Developed an AI football analysis system providing statistical insights on the performance of teams and players. • Fine-tuned YOLO with integration of ByteTrack for consistent detection and tracking of players, referees, and the ball across all video frames. • Improved ball tracking with SAM2, fixing the issue of no ball detection in many frames. • Implemented a team classification system using K-means on jersey pixels; extraction of the two dominant colors (unsupervised approach). • Implemented an Optical Flow and Feature Detection algorithm to estimate and compensate for camera movements. • Developed perspective transformation to convert pixel coordinates into real-world coordinates. • Integrated the calculation of players’ speed (km/h) and cumulative distance (m) per player. • Designed an algorithm for assigning the ball to players in possession and generating ball possession statistics for both teams. • Developed the complete visualization system with bounding boxes for the players, triangular marker for the player with the ball, ellipses for the referees, and statistics overlay. -Technologies: Python, YOLO, supervision, OpenCV, scikit-learn, NumPy, Pandas, Matplotlib

• Developed a flight ticket price prediction system to optimize pricing and personalization of travel packages. • Developed a complete data preprocessing pipeline using Python, Pandas, and NumPy: cleaned irrelevant columns, removed outliers, preserved the last occurrence of duplicate rows, and filtered out insufficiently documented routes. • Integrated a feature engineering strategy: extracted temporal components (day, hour, minute, flight duration), applied One-Hot Encoding for 12 departure and arrival cities, Label Encoding for the day and class columns, and Min-Max Scaling of time-related numerical features, significantly improving predictive performance. • Implemented and compared three regression models (XGBoost, Random Forest, Extra Trees) using scikit-learn. • Performed multi-criteria evaluation using R², MAE, and RMSE metrics. • XGBoost achieved the best performance: R² = 0.9969, MAE = 4.19, and RMSE = 7.50. • Conducted exploratory data analysis (EDA) with visualization of distributions, correlations, and metrics dashboards via Matplotlib and Seaborn. -Technologies: Python, Pandas, NumPy, Matplotlib, Seaborn, scikit-learn, XGBoost

• Developed Satoripop's Website using Figma designs. • Created 3 versions using Bootstrap/CSS, Sass, and Tailwind.