LLM Chatbot & ASR Pipeline (Tersea)
Production AI system for call centers: an LLM-powered chatbot that assists agents with FAQ retrieval, intent classification, and response generation — paired with a full automatic speech recognition (ASR) pipeline that converts live voice calls into text. Continuously improved ML prediction and recommendation models in production, achieving a +20% performance gain through fine-tuning, cross-validation, and monitoring.
Project Highlights
Designed and shipped an LLM-based agent-assist chatbot used in live call center operations.
Built an end-to-end ASR pipeline integrating Speech-to-Text APIs with custom audio preprocessing.
Improved ML prediction and recommendation models by +20% via fine-tuning and cross-validation.
Established production monitoring for model drift and performance regression.
Explore More Work
Deep dive into other high-performance solutions.
Production AI system for call centers: an LLM-powered chatbot that assists agents with FAQ retrieval, intent classification, and response generation — paired with a full automatic speech recognition (ASR) pipeline that converts live voice calls into text. Continuously improved ML prediction and recommendation models in production, achieving a +20% performance gain through fine-tuning, cross-validation, and monitoring.
Platform that helps individuals and businesses assess the climate and societal risks associated with their properties. I built environmental and societal risk models covering floods, hurricanes, wildfires, earthquakes, air pollution, crime, socioeconomic risk, and health infrastructure — processing large-scale geospatial data and deploying models on cloud infrastructure to deliver property-level risk scores at scale.
End-to-end ML system that predicts power consumption across 3 zones of Tetouan, Morocco. Covers the full lifecycle: data preprocessing, feature engineering, model training and evaluation, experiment tracking with MLflow, containerization with Docker, and production deployment on AWS for scalable predictions exposed via a Flask API.
Two-part operational and decisional system for law enforcement. The operational side handles daily case management; the decisional side uses deep learning for spatial coverage forecasting and decision support. Built as a full-stack web application with Node.js, Express, MySQL, and an ETL layer in Talend feeding TensorFlow / Keras models.
Conversational application that lets users chat with their PDF documents. Built with Langchain for retrieval-augmented generation, Hugging Face for embeddings and LLM inference, and Streamlit for a fast, clean UI. Demonstrates a practical RAG pattern over unstructured documents.
Fraud detection system combining graph database modeling with classical ML. Used Neo4j to model transaction relationships and Random Forest to classify fraudulent patterns, leveraging graph features (centrality, motifs) alongside transactional features.
Streaming data pipeline that ingests real-time stock market data and lands it in a queryable analytics layer. Built on Kafka for streaming, AWS S3 for storage, Glue for cataloging, and Athena for SQL querying — running on EC2.
Web app that automates Instagram posting — generates an image, a quote, and relevant hashtags from a single keyword input. Built with Python, OpenAI for generation, and Flask for the web layer.
Web application for sharing favorite places with a social feed. Full MERN-style stack: Node.js + Express on the API, React on the frontend, MongoDB Atlas for persistence.
Binary classification model predicting whether a borrower will repay their loan. Compared classical ML baselines against a small neural network in TensorFlow / Keras.
Social-media sentiment analysis project using the Twitter API to track opinion around Brexit. Produced descriptive statistics, time-series plots, and word clouds to surface dominant narratives.