Francesco Romandini

The project

The project focused on leveraging geospatial computer vision to develop a geolocated building characteristic extraction service for estimating insurance premiums. Initially, individual microservices were developed, which were later orchestrated into a unified service.

High-level pipeline of the service put into production.

• Handled MAXAR or geospatial government APIs, Google Geocoding and Street View APIs.
• Trained, tested and delivered 4 machine learning models: instance segmentation of building rooftops and solar panels [YOLACT]; object detection of building facade elements (like windows, doors, etc.) [YOLOv3]; image classification of buildings facades to categorize their construction material [Efficient-Net].
• Developed estimation rule-based cv algorithms exploiting ML outputs (e.g., pixel2meter conversion, number of floors estimation, etc.).
• Industrialized in microservices as Lambda functions, orchestrated by Step Functions, supported by S3 and DynamoDB connections for metadata, images or results storage.
• When the service was successfully moved into production as a subscription service for several countries I received very positive feedback as the main contributor in this project.

Improved Skills

• Machine Learning (pytorch, mmdet, YOLACT, YOLO, RtDetrv2)
• Computer Vision (pillow, opencv, shapely)
• Geospatial Data (geopandas, MAXAR APIs, Google APIs, WMS APIs)
• AWS Cloud (S3, Sagemaker, DynamoDB, StepFunctions, Lambda)
• Python (numpy, matplotlib, pandas)

The project

The project aimed to detect tiny anomalies in the Italian electricity transmission grid by analyzing high‑resolution aerial photos captured by helicopters. We developed a solution using a cascade of supervised models: first identifying macro‑components, then micro‑components, which were subsequently analyzed by classification models to detect anomalies.

High-level pipeline of the service put into production.

• Worked in an internal team of 5 data scientist, 3 cloud engineers and 1 PM.
• Developed computer vision algorithms for the pre-annotation of rare anomalies.
• Trained, experimented with, and validated three object detection models and one image classification model (YOLOv5).
• Contributed to building a unified pipeline for processing 50TB cross-account data.
• Supported the team in utilizing AWS effectively, configuring and managing resources such as IAM, EC2, S3, DynamoDB, Sagemaker.

Improved Skills

• Machine Learning (pytorch, YOLO, ultralytics)
• Computer Vision (opencv, shapely)
• AWS Cloud (S3, EC2, DynamoDB, Lambda, IAM)
• Python (numpy, matplotlib, pandas)

The project

The model is framed within a Reinforcement Learning framework (an agent interacting with its environment). The agent is equipped with an artificial neural network model (a LSTM) that is capable of learning the transitions through the environment, internally generating possible sequences of actions, and planning the best path to follow in order to obtain a reward.

Schematic representation of the classical reinforcement learning loop. The environment pass state and reward information to the agent at time t. At the same time the agent performs an action that produce effects on the environment. As a response, the environment outputs another pair of state and reward measures at the next state, and the loop can be repeated. The main functions of theta sequences are highlighted in blue (T1 is the encoding function, T2 is the future path evaluation function). Main functions of SWR sequences are highlighted in green (S1 is the training function of the Bi-LSTM network, S2 is the past rewarded path valuation function).

Improved Skills

• Scientific literature research
• Reinforcement Learning
• Bi-LSTM recurrent neural network
• Object-oriented programming
• Python (numpy, pytorch, matplotlib)

Useful links

• The code can be found in my Github repository.
• The whole model simulates what has been argued to happen in the hippocampal formation of mammal brains. For a review see Olafsdottir et al (2018).
• The cover picture representing "a brain" is offered by rawpixel.com.

The project

Climate change is a serious and urgent problem for humankind. Stakes are high and we need to stop it. A crucial step is to gain awareness of the problem, the next step is to actively work for improving the situation. With this analysis I explored the European Social Survey dataset of 2016/2017 to gather some insights on climate change beliefs of many European countries guided by some questions you have probably already asked yourself.

• Do Europeans actually acknowledge the relevance of climate change?
• Are there some hidden beliefs that might undermine a fast and efficient problem resolution?
• Do European countries have different problems and therefore different directions of improvement?
• Are there apparently unrelated factors that are instead related to climate change beliefs?

Results shows that more than 90% of people acknowledge that the world's climate change is definitely or probably happening. From this group of people not even 45% of them think the climate is changing entirely or mainly due to human activities. Considering just the latter group, slightly more than 50% of them feel a distinct personal responsibility to try to reduce climate change.

Different countries have different response patterns, and therefore different directions of improvement.

Improved Skills

• Exploratory data analysis
• Data Cleaning
• Data Visualization
• Ordered Logistic Regression
• Python (pandas, plotly, scipy, statmodels)

Useful links

• The code can be found in my Github repository.
• The .csv file is from Pascal Biliem.
• The data have been collected by the European Social Survey.
• The cover picture representing "a map" is offered by freepik.com.

The project

I have built a ML pipeline able to classify messages sent during disaster events into meaningful classes (e.g., Food, Medical Help, Shelter, etc...).
During emergency situations a lot of people need the support of disaster relief agencies. However, given the overwhelming number of messages it could be hard to filter the relevant messages and to automatically send the message to the suitable organization. That is why a classification pipeline could be a very useful tool.

The app first provides an overview on the training dataset. Then, when a message in inserted as input, it shows all the classes available and highlights the classes related to the message.

Improved Skills

• Natural Language Processing (preprocessing and feature extraction)
• Machine Learning Pipeline (model fitting and model selection)
• Full-stack web app development (server, database, front-end)
• Python (flask, pandas, plotly, sklearn, sqlalchemy), HTML-JS-CSS (boostrap, jquery), SQL

Useful links

• The code can be found in my Github repository.
• Data were made available by Figure Eight Inc.
• Part of the code was made available by Udacity.
• The cover picture representing "a book" is offered by Joshua Hoehne.

The project

The research line aimed to study the neural activity in the hippocampal formation of epilepsy patients under a verbal fluency task. My contribution was the development of a pipeline of integration, pre-processing and analysis of four types of data: anatomical, electrophysiological, speech, word embeddings.

For each patient four types of data were available. A) Anatomical data such as T1 scans, white matter and hippocampal formation segmentation, and electrodes labels. B) Electrophysiological data coming from an average of 64 channels recording for aproximately 25 minutes. C) Speech data of patients voice recorded while mentioning aloud the highest number of animal names as possible. D) FastText word embeddings freely available online.

Improved Skills

• Data Integration
• Data Cleaning and Preprocessing
• Speech Recognition and Segmentation
• Dimensionality Reduction analysis (PCA)
• Time Frequency analysis (fft, morelet-wavelet)
• Slicing software (Slicer3D)

Useful links

• The project is part of a research line at BottiniLab
• The project is made in collaboration with Ospedale Niguarda .
• The affiliate university is University of Trento.

The project

I have worked with PySpark to analyze a log of user activities interacting with a digital music service. The goal is to predict which users are about to cancel their subscription, aiming to keeping them as customers providing them extra attentions.

I performed data cleaning, exploratory data analysis, feature engineering and modeling. The whole project is very simple and it is supposed to be an excuse to exercise with pySpark.

Besides that, the modeling section has reserved a nice problem to face:

The model is biased toward predicting users as active almost every time, leading to a big amount of false negatives and an overall weak prediction performance.

This is due to the class imbalance problem. In the dataset the number of active users were 2.5x the number of churn users. This is an intriguing problem to solve, yet some methods exists to mitigate the problem (downsampling / upsampling tecniques).

The aim of this work was to familiarize with PySpark logic and syntax. However, there are a lot of improvements that can start from here and take this project to another level: addressing the class imbalance problem more seriously, using a bigger dataset, leaning against a cloud computing service to exploit all the power of spark with big data, building a web application to deploy the model with a comfortable user interface, and much more.

I am currently diving into AWS. I'm planning to come back to this project in the near future to perform some of the improvements mentioned above.

Improved Skills

• ETL
• Feature Engineering
• Python (PySkark)

Useful links

• The code can be found in my Github repository.
• The cover picture representing "a network" is offered by rawpixel.com

The project

I have analyzed the interactions between users and articles of the IBM Watson Studio platform. The goal was to build a recommendation system: to suggest users some articles they might like and they have not seen already.

The Dataset was provided by IBM and manily consisted in a log of interactions between users and articles. There are many ways to provide recommendations, and I took this project as an exercise to try serveral ones on the same data.

• Rank-Based: recommend the most popular articles.
• User-Based Collaborative Filtering: recommend articles popular among similar users.
• Matrix Factorization: decompose a user-article matrix and extract the latent factors that help to predict which are the best recommendations.

Improved Skills

• ETL
• Recommendation Systems
• Python (pandas, numpy)

Useful links

• The code can be found in my Github repository.
• The cover picture representing "5 stars" is offered by DilokaStudio