Why AI is so important @MrJeff?

By Carlos Ruiz

I ended up last year with a tweet commenting about some news related to Mr Jeff: Andrew Ng founded Landing.ai, a company aimed at helping traditional businesses (i.e. manufacturing sector) transform themselves into AI companies. For example, one product that Landing.ai has developed is a visual inspection system that uses images from a camera to identify defects in products. You can see Andrew’s presentation about it here.

My comment was: “I’ll point to this to those who still do not understand why #ArtificialIntelligence is important @MrJeffApp”, but I did not give more insights about why or how. Well, here it comes:.

It might sound somehow a stereotype but it is true: At Mr Jeff, we are re-imagining the very-traditional laundry industry by applying Technology and Artificial Intelligence. Don’t get me wrong, just implementing some AI techniques does not make a technology company

<aside note: yes, Mr Jeff is a tech company, applying constant technology innovation for enhancing customer experience, optimizing operations, pick-up and delivery time, and finally, improving the conversion as much as possible>

in an AI company, but it has all the potential to transform the industry as we expect.

Offtopic: due to my background in both Technology and Artificial Intelligence, plenty people asked me about changing my field when I moved to Mr Jeff, but it happens not to be such a change.

These are some of the things what we do @MrJeff:

Optimizing operations

Mr Jeff is a very heavy-based operations company. In a simplified manner: a customer books an order via app scheduling pick-up and delivery time in a particular address, the operations team set up all the logistic to collect it and drop clothes at the laundering point, the cleaning facility team tags and video records the clothing which goes for washing, dry cleaning, folding, ironing, and packing, to come back to the operations team to deliver the order on the scheduled date. Not even talking about changes or incidents. Uh, tough enough!

All this process requires things like driver’s demand prediction, availability calculation, and route optimization to address customers’ expectation and increase the ratio orders per hour. Different improvements has been included for the last year to automate and optimize all the single steps.

Given an example: how the daily routes are managed.

Known as the Vehicle Routing Problem, this is a classical problems from 60s in Artificial Intelligence, whose goal is to find the optimal set of routes for a fleet of vehicles delivering services to various locations given a number of constraints (time windows, vehicule capacity, multiple depots …). Determining the optimal solution is NP-Hard, so the size of the problems that can be solved using combinatorial optimization may be limited within reasonable computing when the number of orders is large. However, metaheuristics as a way to provide a sufficiently good solution to a large and complex optimization problem is proven to help out.

When we started two years ago, the process was manual and decoupled from the back-office. Few months later, operations was overwhelmed by the numbers of orders they had to manually fit into more and more complex routes. Today it is fully supported by our back-office which provides a metaheuristic-based solution in an automatic fashion. Not ended yet, we are adding more contextual factors about our customers.

Simplifying the interaction with customers

A key part of the success is ensuring the wow experience of the service. Among other things, it implies simplifying the interaction by leveraging the customer relationship with additional means. Although chatbots in channels such as Facebook are becoming more and more popular, an in-app chatbot can increase customer retention and decrease loads on call centers while the customer remains in our platform.

While not deployed yet, at Mr Jeff Engineering we developed a prototype providing general information about products, additional pre-actions within our e-commerce (e.g. place a new order, change delivery dates), but also responding simple questions (e.g. billing) or trigger additional actions (e.g. requesting a human operator). It is worth noting that, instead of asking for each piece of information about the customer or the order in a step-by-step approach, the system retrieves all related information and applies some intelligence to shortcut a larger interaction path. Then, customers can change and order or raise an issue in a more natural way.

Finally, due to the increasing size of the customer base, an additional benefit of such enhancements in customer contact automation capability is to further reduce demand on customer care staff and operational costs.

Understanding our customers: Mr Jeff Camera

Quite often any startup claims how important is understanding their customers, how they purchase, how they behave. Most of them analyse their transactional data to discover insights and trends about purchases, products, and customers’ behaviour. That is something we also do to modify our products or improve our operations, but there is something else we will be focusing on this year.

There are several important, but time-consuming tasks in a laundry and dry-cleaning sector which are currently done manually. These range from validation of the incoming orders (compare order with actual clothing items) to quality control of the items (check for remaining wrinkles in the ironed clothing) before packaging and delivery. Furthermore, there is valuable information about the customer preferences (type of clothing, brands, …) which would be available throughout the process, but is currently not exploited as it would add time and costs to the process. In recent years, with the breakthroughs in artificial intelligence and computer vision with deep neural networks for many challenging tasks (like recognition of handwritten digits or traffic signs) a machine learning algorithm is now able to carry out some of the tasks and beat a human. Specifically, for fashion item classification also several datasets, and promising algorithms (focusing an photos from social media and catalogues) have appeared.

Since we already record all the orders for operational purposes, we are working on a prototype of the Mr Jeff Camera, which is capable of automating some of the mentioned tasks at human level. Although we will be training our models from scratch in near future, at this point we rely on transfer learning over some of the available MobileNets models, the family of a family of mobile-first computer vision models for TensorFlow, designed to effectively maximize accuracy while being mindful of the restricted resources for an on-device. It is only the beginning, since there are some other features still out of the scope (e.g. wrinkles).