Be the first to read our first Ebook!

In the Age of AutoML, we are all Data Scientists! | Artificial Intelligence (AI) has become increasingly essential to our daily routine lives.

SmartPredict users provided insights on their experience

"SmartPredict has definitely helped me manage my most impatient and demanding customers, which is a real relief for me and my colleagues"

D.FI Ticket Referral


"I was bored every time I had to look for a single piece of mail hidden in a pile of unrelated mail. Now I am serene to receive my mail every morning: SmartPredict has definitely helped me manage my most impatient and demanding customers, which is a real relief for me and my colleagues".

D.FI and SmartPredict have already collaborated on many AI-based projects, and this time the project is to speed up the sorting of emails so that they can be processed in a timely manner.

Problem statement and context

D.FI : parsing e-mail complaints and redirect e-mails to the technical team's correspondent, and detect issues similarities to propose a resolution of the problem automatically.

Emails are means of correspondence between the  stakeholders in a project. However, with many emails entering and issuing from mail boxes, despite labels , it is hard to pinpoint those where clients report issues that need urgent troubleshooting or express their discontentment  over a product or service. 

Furthermore, tension may escalate even more from these latter with a delayed processing of requests or presumptions that they are being ignored.

"Sometimes I wish I could sleep at night thinking that everyone is well taken care of, but then I get an incredibly angry letter, when only then do I notice that it is the third or even fourth in a list of previous complaints about the same misadventures", says an engineer at the department of maintenance.

In the meantime, it is hard for the service of processing as a mere human being to sort this huge amount of mails both visually and manually and afterwards, to route the whole to the corresponding responsible.

To solve this problem, the company requires Robotic Process Automation as a means to alleviate this repetitive and strenuous task. 

This will not only allow time to be spent on higher value-added tasks, but will also prevent the company from customer churns, which, if they were to occur, would have a terrible impact on revenues…

In addition, the underlying values of process automation lie in the ability to eliminate human error caused by the unpleasant repetition of the same actions over and over again. By focusing more  on other tasks such as solving the issue at hand, one can immediately determine which  action to undertake strictly speaking.

In fact, what used  to be a monotonous task for a human, is just a piece of cake for Artificial Intelligence.

Thanks to  SmartPredict’s fast classification algorithms, users simply have to connect samples of previous mails for the engine to analyze and categorize according to the topic and subject of the mails . Afterwards, the ML- backed chooses the best fitting technical support and redirects the mail to this person.


The first challenge was the lack of data. However, through skillful scaling of measurements, SmartPredict was able to manage the optimization of proportions for training and validation sets. The usual flow of scientific data pipelines is still being modernized which allows room for sophistication in the mechanics of processing, while providing continuously a high level of simplicity to the end-user.The model is not static, but of course it is improving over time.

With a sufficient data set, the model performs well above 95%.


Using a Serverless Machine Learning model, bundles of a thousand mails are analyzed and sorted out to extract useful information then redirected to the compounding addressee.  

An interface for Front-end is built to ingest the dataset made of those  mails and then be parsed through Machine Learning classifiers .Finally  a Convolutional Neural Network will be trained to identify the mails.

 Starting from a scraping and parsing of text through Natural Language Processing, SmartPredict incorporates specialized Machine Learning algorithms to gain accuracy in the predictions.

  • Regression and classification models are built using TensorFlow,  then these models will be deployed on AI Platform so that Cloud Functions can use them later.
  • Machine learning models: 
  • Support Vector Machine (SVM), Classifier
  • Extreme Gradient Boosting (XGBoost)
  • Random Forest Classifier
  • Convolutional Neural Networks:
  • 1D CNN
  •  Similarity Algorithm:
        -       Cosine Distance


The results were positive overall and ultimately brought excellent and sustainable results for both the company and its customers.

“We are pleased that now, less and less customer reports are received. Thanks to SmartPredict , emails are indexed according to the exact subject of complaints which is valuable to address the issues at the core,while maintaining a positive relationship between us and our clients- improving their overall lifetime value.  ”

Applications are now processed in a timely manner and are no longer subject to queues as they are automatically assigned to the right person.

D.Fi's routing tasks are now managed seamlessly, resulting in better satisfaction for the processing of data.

This sums up the experience D.FI had with SmartPredict which is in a nutshell,simply a positive smart enterprise.


We get some positive feedback from you

"I'm excited to see what SmartPredict can provide. So far, I'm thoroughly impressed."


Software Engineer

“I have tested Autoflow and I think it is quite innovative because it offers the possibility to customize the generated flowchart, which I don't think I've seen before. It also allowed me to reduce the development time to a few hours for the forecasting project I've been working on for days".

Marius K.

Business Manager

"I'm really glad I found this tool and wondered why I didn't use it earlier. At first, I was a bit skeptical and thought that getting to such a level of abstraction was near impossible, as was making the algorithm connections with modules. The only thing that demanded a little more attention was to identify the sequence of modules from code to flowchart, but once you get used to it, it seems really simple."

John P.

Expert Data Scientist

Stay up to date!

This email has already subscribed.