Hadil Sahraoui

• Achieved 90% recall in churn prediction by engineering XGBoost model with RFM features on 500K+ subscriber

records, resulting in identification of 12,000+ at-risk customers monthly and 85% reduction in manual segmentation

effort for the analytics team

• Achieved 0.85 Silhouette Score in customer segmentation by implementing K-Means clustering on 100K+ users with

behavioral features, resulting in 5 distinct segments that enabled targeted campaign design during marketing pilot testing

• Accelerated content generation 15x by building LangChain + Qwen 2.5 LLM pipeline with dynamic prompt templates

for 50K+ subscribers, resulting in reduction from 40 to <3 hours weekly manual effort and freeing 15+ analyst hours

weekly for strategic work

• Reduced query resolution time 87% by engineering RAG pipeline with ChromaDB vector store and Qwen 2.5 LLM to

answer 200+ daily internal queries in <2 minutes , resulting in adoption by 5+ support agents during internship pilot

• Enabled production-ready deployment by containerizing Django prediction API with Docker and deploying on Microsoft

Azure, resulting in 99.5% uptime during load testing with 5K+ simulated daily requests and seamless integration with

internal analytics tools

Developed a Cheque Autocorrection System using OCR, machine learning, and NLP. Integrated KNN for digit extraction and PaddleOCR for text recognition.

Deployed a full-stack Django web app with user authentication and utilized PowerBI for visual analytics.

Collected and processed data via APIs and web scraping to enhance dataset quality. Fine-tuned Large Language Models (LLMs) for NLP tasks and performed data analysis to extract insights for decision-making.