Data Scientist and Developer in Reggio Emilia, Province of Reggio Emilia, Italy

The need was to create a data-agnostic cloud-based platform for ML experiments and production lifecycles, decoupled from the main business analytics software.

I designed and implemented an Azure-based distributed platform, which included (1) Serverless Azure functions as the platform API and workflow orchestrator, (2) Azure Blob Storage as the datalake, (3) MSSQL DB (later moved to Azure Tables for convenience) for storing states and intermediate results, (4) Azure ML Compute Clusters for running ML algorithms and producing artifacts, (5) Azure Kubernetes Cluster for deploying the models and serving the predictions, (6) Git repository of algorithms which can be run on-demand, (7) CI/CD pipelines.

When a user creates a project and uploads the dataset to the main software, the platform accesses the data and runs a series of ML experiments, producing models, predictions and explanations. The predictions and explanations are submitted through the REST API back to the main software, where they are displayed to the user in various scenarios, providing them with detailed insight into their dataset and allowing better decision making. Some of the models are automatically deployed as endpoints to provide real-time predictions.

Several stages of the drug manufacturing at the client's production plant had a lead time of 1.5 hours per batch and needed constant manual intervention, which prevented the desired scaling of the production. The client and I identified the root cause as a lack of a precise model of the amount of chemicals needed to add to each batch to achieve desired product properties.

Using historic process data obtained from the client, I created a precise model, which allowed me to combine several production steps without the direct involvement of the personnel. I achieved this by extracting necessary variables from the time-series signals of the production line sensors and combining them in polynomial regression. This reduced the lead time of the bottleneck manufacturing step to slightly less than 30 minutes, leading to a corresponding increase in the throughput (+200% as estimated by the client) while also reducing the load on the technical staff. I packaged and shipped the model, meeting specific client technical requirements.

While working on the project, I proposed several cost-effective improvements to the production line, which are estimated to reduce manual involvement by 200% while increasing the production's precision.

The goal:
- Extract business rules describing the conditions under which a process goes from activity A to one of the possible next activities (B, C, etc.)
- Estimate the consistency of these rules.
- Present this in a user-friendly form.
- Take <5 seconds on 1 million business cases.
- Integrate easily with the main Java software.

I decided to use Java 8 for that project. I extended the publicly available basic Decision Tree library with many necessary functions, such as pruning, metric estimation, tracking groups, working with missing data, and more. With that and feature engineering/augmentation pipelines, the algorithm obtains classification models for each transition and translates them to text rules, such as "A to B: when X > 10, or X < 1 and Y > 100". These are easily interpreted by business users. Metrics are presented in a user-friendly way, allowing to judge the consistency of the identified rules. I designed UX/UI for displaying and exploring the insights and adjusting them to the specific users' datasets (e.g., users can make rules more complex and precise, if they want).

BRM became one of the core features of the software anda key sales point. It became a basis for process simulations, another core feature.

We created and tested a brain-computer interface prototype based on the real-time analysis of multidimensional bioelectrical signals obtained from the scalp of a car driver (EEG), showing the selected items in the infotainment menu in a completely hands-free way.

I designed and implemented an EEG data preprocessing pipeline and ML model (based on the convolutional neural network architecture), which was trained on the driver's data and in real-time predicted which infotainment function they wanted to select (navigation, music, etc.). The ML model was ported to the battery-powered portable edge device NVIDIA Jetson TX2, allowing it to work independently inside a car. To increase the project business value, we also collected rich motion data using a set of accelerometer sensors to create future models predicting steering actions.

I co-led this project; in particular, I coordinated the activity between neuroscientists, engineers, and our research partners from Toyota Motor Europe, designed the prototype tests and data collection.

This work resulted in several papers (for a detailed account, see the project URL) and a patent I co-authored (https://patents.google.com/patent/WO2020211958A1).

I analyzed the electroencephalographic (EEG) dataset to extract patterns related to the driving actions; braking, acceleration, and steering. The data consisted of an EEG, accelerometer data, and driving simulator data, all of which were multidimensional time series.

I was invited to the project at a late stage; however, while analyzing the previous work, I found a serious bug in the data analysis, which invalidated the results about to be published. As a consequence, I was asked to join the project full-time to improve the analysis, which was eventually published as a scientific paper and partially patented (https://patents.google.com/patent/WO2019025000A1).

In particular, my work consisted of synchronization of the data streams from different devices, extracting and filtering event triggers, performing PCA, and factorizing EEG signals on independent components (ICA), with subsequent time-frequency statistical analysis.