Home Blog Our relevant cases o ...

Our relevant cases of Machine Learning Model Serving

Marco Belmondo

We make difficult things easy by modelling data through Machine Learning

Intelligently collecting data and modelling it through AI algorithms enables companies to solve business problems
or open up new opportunities. Here are some of our work that may inspire you.

AI for predictive maintenance


Requirement: reduce production system downtime

Activities: starting from the existing network of sensors, with the
possibility of introducing new ones (temperature, accelerometer/vibration, current clamp,
cameras) in specific areas of interest, we created a unitary infrastructure (datalake, data
pipeline) able to store data in quasi-real-time.
Subsequently, leveraging the collected data, several Machine Learning algorithms were
trained to monitor the KPIs of interest and recognize any anomalies.
Finally, a web dashboard was created to describe the status of the entire supply chain and
highlight any anomalies and criticalities.

Output: through the implementation of different solutions, able to control
at 360° the sensors of the entire production chain, we intervened in the monitoring of all
rotating devices (motors, pumps, extruders, conveyor belts, etc.), providing the customer
with a dashboard of easy consultation.
In both Food and Plastics sectors, we now have a strong experience in monitoring for
predictive-maintenance purposes of the machines involved in the whole production process.

Results: In just 6 months, we started with the requirements gathering and
then completed the data modeling study, bringing into production an AI-based solution that
made the difference. One year after the production deployment, the results are very

  • 3 million Euros saved thanks to “early-warning” on anomalies and material damage,
  • 30% reduction in maintenance costs
  • 25% increase in production
  • 25% reduction in downtime

Predictive AI for churn decrease



Predictively intercept customers at risk of abandonment in order to carry out
personalised marketing activities.

Using a machine learning model, we analyse the customer’s interaction path with the company.
calculation of a risk index relative to the abandonment (churn rate) of each customer for
each specific service.

In detail, for each customer we periodically calculate his churn score based on all available
data that allow us to reconstruct his behaviour with the company (web tracking, crm, email,
transactions) thus defining clusters of users by risk class (high risk, average and power


  • creation of a risk index for each customer
  • creation of clusters of users at high risk of churn

– 14.7% churn rate after only 2.5 months of activity

Predictive AI for customer segmentation (case 1)


Requirement 1:

Increase the number of users who complete their first money deposit (the “prospects”)

Requirement 2:

Intercept quality users. a “quality user” is defined as a user who continues to deposit
and use money in games over time.

Activity 1:
– management of data flows (web analytics, adv, crm) from different teams operating with
separate silos (data lake)
– creation of a “path to registration” that identifies all user touchpoints, from arrival at
the site to registration.
Activity 2:
– user segmentation using clustering algorithms that work on metrics such as frequency of
play, volume of deposits, specific interests, browsing times. the algorithm also uses
insights from the path to conversion to increase the quality score built on the clustering.

– development of a user quality index (similarity index)
– customisation of the offer based on the identified segments.

+18.7% conversion rate
The identification and visualisation of the “quality user” cluster led to an increase in the
conversion rate from registration to new deposit.

Predictive AI for customer segmentation (case 2)



To get the most out of internal proprietary data (first-party data) in order to acquire
new customers and retain existing ones.

– integration of online navigation data, crm, email marketing and other traffic sources into
a single environment.
– use of proprietary predictive algorithms to segment italotreno business clients by
“lifetime value” (re-engage client, high potential, frequent client, top client).
– use of proprietary predictive algorithms to associate a probability of repurchase to each

Creation of more profitable customer segments to which to address both email marketing
activities and advertising investments (e.g. remarketing).

+20% conversion rate from remarketing activities, reducing investment by

+3% conversion rate from prospecting activities,

Read more in this

Predictive AI for sales forecasts

Media & Telecommunication


To forecast sales volumes related to the launch of a new product in the telco sector in
order to assess its impact on total revenues.

Creation of a predictive model based on neural networks capable of accurately estimating the
sales volumes of the new product both pre (*) and post launch.

+15% prediction reliability

(*) model trained with a 2-year history of similar product sales and enriched with a
dataset including quantifiers of advertising pressure (budget spent and planned for
advertising diversified by channels, tv grps or equivalent), of specific promotions and
with the average price of equivalent products available in the market.

Forecasting credit scoring

Risk e Credit Management


To improve the models through which the default risk of unlisted companies is measured.

Apply machine-learning algorithms to the analysis of financial data from chambers of
commerce to predict default.
create a digital identity score of each company (online presence, activity and reputation)
by enriching the available datasets.
integrate the financial signal with digital information to refine the prediction of default
risk compared to traditional methods.

+13% increase in prediction reliability (from 80% to 93%) by applying the
new scoring model to a sample of 135,000 smes using over 1,400 variables.

New business prediction



To increase the opening of new current accounts.

Analyse all user touchpoints, from arrival at the site to account opening.
identify which sections and categories of the site users visit most often before opening an
account (path to conversion) – e.g. faq section.
identify the behaviour patterns with the highest probability of conversion (clustering,
behaviour prediction and interest graph models).

+34% opening of new accounts

Prediction of product launches in physical stores

Food & Beverage


To predict the potential and/or effectiveness of cross-country physical outlets by
identifying potential shopper pools.

To characterise the real audience of one or more outlets within specific markets by
combining information linked to territorial socio-economic data, characteristics of
distribution chains, alternative digital signals.
creation of an alignment score of the outlets with respect to the target outlets.
indication and evaluation of new product launches in specific outlets in the american and
chinese markets.

+32% revenue vs forecast

B2B lead generation (case 1)



To increase the database of new target companies

Using supervised computer vision and machine learning algorithms, we created a model capable
of analysing google maps images to identify buildings having roofs with characteristics and
surfaces suitable for the installation of photovoltaic systems. the analysis was performed
on geographical areas of interest indicated by the client.
subsequently, all available information from the google my business quadrant was extracted
to identify the companies and extract the available information.

Sharing of a company database with master data and business contact priorities (address +
contacts, geolocation, photos of the property and estimated roof area, product propensity
The client then implemented customised marketing and sales activities based on the company
database obtained from the information collected.

B2B lead generation (case 2)



Lead generation through ai and targeted advertising campaigns

Collection and processing of satellite images (google maps and open street maps)
of a
specific italian coastal region (2 km from the sea) using the most advanced ai
and computer
vision techniques to identify roofs with a surface area greater than 350 m2.
construction of micro-targets in a capillary way thanks to the evidence from the
point and other external data sources (e.g. job title linkedin).
planning and launching of advertising campaigns in the specific areas of
interest (google
search, google display, youtube and facebook). for each lead a conversion
propensity score
is given.