Francesco Fontan

I led a team of three people to design a machine learning tool on Microsoft Azure to improve the ticketing management of a telco customer.

This system analyzed emails received from customer support and logs from the network infrastructure, extracting useful information using NLP techniques and LLMs to open, assign, and finally close tickets fully automatically.

This automation increased the speed by ten times, predicting more than ten ticket fields with an average accuracy of 90%. Finally, the pipeline could scale: training was performed on half a million records, while the inference module handled more than 20 tickets per hour.

Implementing predictive maintenance for highway asphalt is a pivotal approach to cost reduction and enhancing the overall service quality for motorists. Within this project, I engineered a sophisticated system to identify and categorize diverse road pavement defects effectively.

This solution harnesses an object detection model, utilizing Yolo v5, in conjunction with external data points pertinent to the road section under scrutiny, such as traffic patterns and weather conditions.

Notably, the recall rate surpassed 80%, and the tool achieved significant cost savings of 600,000 euros in its inaugural year of operation.

I created a versatile chat and voice bot for Question and Answer (QA) and Requesting Assistance and Guidance (RAG) using Llama 2, LangChain, and Weaviate. This system seamlessly integrates chat and voice features, including phone call interactions using Twilio, speech-to-text (Whisper AI), and text-to-speech (Polly) models. The bot's Natural Language to SQL module enhances information retrieval and user interactions, effectively managing approximately 1,000 daily calls with an 85% success rate. Implemented in Python with FastAPI, the system offers robust performance.

The main goal was to use artificial intelligence to enhance the planning phase of crew members by optimizing the global fleet planning by minimizing the money spent on flight tickets without lowering the expected quality.

I designed a Python tool based on Google OR-Tools, an optimization suite by Google AI, able to solve MIP problems that involved more than 30,000 embarks every year, obtaining an estimated average saving of 10%, corresponding to $1 million per year.

In my role at Machine Learning Reply, I have been instrumental in offering a range of training courses tailored to meet the unique needs of our clientele. My primary focus involves delivering workshops that delve into the intricate realms of deep learning and accelerated computing.

Courses Delivered:

Fundamentals of Deep Learning
Building Transformer-Based Natural Language Processing Applications
Building Intelligent Recommender Systems
Fundamentals of Accelerated Data Science with RAPIDS
Fundamentals of Deep Learning for Multiple Data Types
Fundamentals of Accelerated Computing with CUDA Python and C/C++

As the technical team leader for this project at an airline company, I spearheaded the development of a tool to optimize budget allocation. The objective was to recommend the most effective distribution of the marketing budget between campaigns to enhance traffic volume and conversion rates.

To achieve this, we crafted forecasting models that meticulously analyzed ticket sales over time, factoring in external variables such as weather, competitor actions, and specific events. Additionally, we engineered a system capable of predicting future visitor numbers based on a set of marketing campaigns.

Through these efforts, we estimated an annual cost savings of approximately $300,000.

I led the overhaul of a critical machine learning model and its associated data pipeline, which played a pivotal role in Swaption price prediction by applying the Black-Scholes theorem. My efforts yielded significant improvements, most notably a remarkable 10% reduction in Mean Squared Error (MSE) compared to the previous implementation.

Additionally, I achieved resource optimization by slashing RAM requirements by an impressive 30% while accelerating model processing by a factor of 1.5x. To fortify the model's reliability, I strategically employed Ensemble methods, harnessing the strengths of XGBoost, Random Forest, and Neural Network techniques.

I developed a recommender system for Reply's internal social network using traditional collaborative filtering methods, item-based models, but also NLP techniques to extrapolate information from posts and user profiles. These algorithms were entirely on the AWS cloud and handled more than 10k
active daily users.

For an airline company, two metrics are the key to success: conversion rate and traffic volume.

I was the technical team leader in this project that aimed to define a tool capable of suggesting how to distribute, in the most effective way, the budget between marketing campaigns to increase volumes and more aggressive pricing policies to increase the conversion rate.

We developed forecasting models to analyze the tickets sold over time and consider external factors such as weather, competitors, and particular events. But also a system able to predict the number of visitors in the future given a set of marketing campaigns. Finally, we estimated a cost saving of around $300,000 per year.

A versatile chatbot and voicebot that I designed for Q&A and requesting assistance and guidance. This system seamlessly integrates chat and voice features, including phone call interactions through technologies like Twilio, speech-to-text, and text-to-speech models. The bot's natural language to SQL module enhances information retrieval and user interactions, effectively managing approximately 1,000 daily calls with an 85% success rate. Implemented in Python with FastAPI, the system offers robust performance.

As the project leader, I spearheaded the automation of more than 5,000 product description generations in over 20 languages using Generative AI. Within a team of five, we successfully developed a promising prototype and are currently working on scaling the solution for multiple markets. Our methodology involves leveraging product attributes, images, and customer reviews to create HTML descriptions for the website, employing a pipeline with three language model modules and a focus on prompt engineering and few-shot learning using ImageGen and Palm2, with potential plans for fine-tuning in the future.

I spearheaded a dynamic pricing optimization initiative at Levi's, overseeing a team of four specialists across data science and engineering domains. Our project targeted promotional periods across Europe, the US, and Mexico, covering mainline, outlet, and eCommerce channels. Through demand forecasting utilizing sktime, statsforecast, and XGBoost, coupled with optimization via mixed integer programming, we maximized quantity sold, revenue, and margin while adhering to inventory constraints and business rules. Deployment on GCP with Kubeflow and Airflow ensured scalability, supported by CI/CD with Jenkins. This implementation contributed over $40 million in annual revenue, affirming the significant impact of our data-driven strategies on financial outcomes.

Led a bundle promotion optimization project aimed at enhancing markdown and pricing strategies by incorporating bundle offers such as "Buy one, get another item at 50%." Utilizing demand forecasting techniques with sktime, statsforecast, and XGBoost, we integrated mixed integer programming optimization to maximize quantity sold, revenue, and margin while considering inventory constraints and business rules. Deployment in US outlets led to a remarkable revenue increase of over 10% based on recent A/B tests facilitated by Kubeflow for scalability and effectiveness.