Quelques réalisations
Example 1 - Smart Delivery Estimate
Shipup is the French leader in post-purchase experience. By creating a reassuring, intuitive and customer-focused post-purchase experience, companies can turn one-time buyers into regular customers and small e-tailers can compete with the giants on the quality of service.
Shipup asked Scopeo to create a Machine Learning model to estimate the likely delivery date of a parcel at the time of ordering and then throughout its transport.
To be able to answer this problem, which is an example of predictive analysis, we followed these steps:
Example 2 - Automated document parsing
Un de nos clients développe un logiciel d’aide à la comptabilité d’entreprises. Il a commandé à Scopeo to develop an algorithm for automatically reading scanned invoices..
To meet this demand, we have divided the analysis pipeline into different stages:
OCR analysis
Tagging et contextualisation
Features extraction
Disambiguation
This allowed us to develop an easy-to-use software library.
Original document
Initially, the image is obtained by scanning the document. It contains text but also other types of print such as table lines, logos, stamps, signatures etc.
Step 1: OCR analysis
We apply state-of-the-art technologies to read the text and identify its position on the document.
Step 2 : Tagging and context
We interpret this text using Natural Language Processing techniques, taking into account tables, rows, context...
Step 3 : Features extraction
Then, we add all the features that characterize this document type and the fields to extract, in order to give to the algorithm as much information as possible.
Step 4: Disambiguation
Finally, we implement rules using business knowledge on the one hand and statistical observations on the other (Machine Learning) to assign the corresponding content to each searched field.
Example 3 - Credit default prediction
Fydem, a Credit as a Payment specialist, called on Scopeo's services to develop a tool that could categorise transactions from Open Banking (PSD2) data and quantify a borrower's risk of default or estimate the NBI (Net Banking Income) of a potential customer.
We supported Fydem in several ways:
Support for finding and aggregating relevant data sources.
Development of production code and structuring of engine for internal use
Construction of prediction algorithms
Support for recruitment (tests, interviews), training and coaching of recruits
Example 4 - Customer reviews analysis
In order to have some feedback from customers on a product, it is interesting to study online reviews. Thanks to this source of information, you can iterate on a product using customer impressions. This being said, it can be overwhelming each product have tens of thousands of reviews.
To solve this, we developped, for our client, a tool that synthesize all comments on a product.
Our tool uses state-of-the-art Natural Language Processing algorithms to isolate sentences dealing with the product and summarizing them all.
The use of very powerful pre-trained neural networks allows us to get results with very little labeled data.
Example 5 - Classification and recommendation of fonts
In order to have some feedback from customers on a product, it is interesting to study online reviews. Thanks to this source of information, you can iterate on a product using customer impressions. This being said, it can be overwhelming each product have tens of thousands of reviews.
To solve this, we developped, for our client, a tool that synthesize all comments on a product.
Our tool uses state-of-the-art Natural Language Processing algorithms to isolate sentences dealing with the product and summarizing them all.
The use of very powerful pre-trained neural networks allows us to get results with very little labeled data.
Fonts Ninja redefine the market of typographies as they help foundries to control their licences and the designers to discover and buy new typographies.
The client wanted to set up a tool to look for and discover typographies easily and efficiently and to find all typographies similar to a given one. The difficulty here comes from the subtility of the differences between typo styles.
We trained deep learning algorithms to both classify and project images into a vectorial space, allowing us many other tasks such as recommendation. Everything is integrated into a micro-service and do not require any Machine Learning Expertise to be used.
We had before a user experience where one couldn't find similar fonts and had to scroll down and look, to find a font that matches.
We have now a suggested list of fonts that are similar to a given one. Fonts are now classed based on many criteria and allow the user to search for some specific style of typographies.
