Building the future of AI-assisted peer-to-peer marketplaces at vendi

Photo by Andy Kelly on Unsplash

The need to buy or sell is fundamental and is not new, it has always been part of the human nature. It has always revolved around some transaction that is of value to parties involved. This phenomenon has lead to marketplaces of all kinds.

With the onset of the internet revolution , it’s only natural that these marketplaces shifted to be online. The evolution of such marketplaces are covered in plenty of blog posts like this.

Marketplaces used to be governed by buyers being physically limited to purchasing one or two things at a time. For example, the age old barter systems where you sell something if you have to buy something. Nowadays, this has changed and with the presence of online marketplaces buyer can purchase as many items as they want.

What this means is that many of us just buy too many things but never understand the value of said items fully when they are no longer of use to us. If this seems unreal, just try and remember the last time you sold something. Now think about all that you own which is in good condition but you didn’t bother to discard, recycle or sell and make money with. I am sure some of us even have items that have not been opened, like accessories for devices. I, for instance, have several power chords/ headphones and others lying around.

So why is it that we are so good at buying but not selling. A major factor is convenience. Buying has become quick, instantly clickable, and easy. On the other hand, selling still requires thought, writing descriptions, pricing, managing logistics… With all the advances in internet technologies, information, and online security this side of the marketplace is still not enticing or convenient enough. I have covered this in a previous blog post.

So how to improve this and what is next? We have seen some positive changes in the peer-to-peer (P2P) marketplaces over time and companies like vendi are passionately disrupting the way we buy and sell in P2P marketplaces. One particular tool that is of significant importance in this change and is touted as the herald of the next industrial revolution is Artificial Intelligence (AI) and machine learning (ML) driven engines. So let’s talk about how AI is going to change P2P marketplaces. For the rest of the blog we will call this AI-assisted P2P marketplaces.

Before we jump right in let’s set some expectations on what is AI. There has been a lot of hype for good reason. But clarifying reality from hype would help set some expectations. I am also inclined to not propose something that is not possible today or at least working in production in some data centers.

What are the tools of AI that can assist a P2P marketplace. We will look into two major categories where Deep Learning has made tremendous progress in the past view years.

Vision

Vision is where Deep Learning (DL) techniques have exponentially improved results for object recognition ever since the ImageNet competition in 2012 when DL took over existing computer vision technologies. It has come a long way since and many architectures have continuously pushed the state of the art.

With the advent of Generative Adversarial networks (GANS) there has been a lot of progress in style transfer, AI art, super resolution and coloring.

One of the major benefits of object recognition in the marketplace is visual search. The paper DeViSE covered here, links ImageNet to word embeddings and showed impressive combinations of tooling with both word embedding and image feature maps the technology has come a long way. There are many products that use visual search in some way or the other.

Text

Text is usually attributed as Natural Language Processing (NLP), though there is a lot of debate over how natural AI is and whether language intelligence as we know is solved by AI. What is clear is that it’s getting better. Some have even called last year as the ImageNet moment in NLP. There has been lot of progress in speech recognition where platforms like Alexa, Google and Cortana have become mainstream. Though there are still some misses, there are far more hits on daily usage. Google famously demoed and has debuted for limited audience its Duplex feature. So it is not an exaggeration to say that bots as personal assistants are not very unreal, though there are limitations to them.

Having set the expectations on the current state of these tools, let me cover how we use and plan to use AI for assistance in a P2P marketplace at vendi.

Let’s cover the use of AI at vendi in 3 categories:

AI assistance for: SELLING

We had covered how easy it is to buy vs sell as individuals online. Big online retailers who sell have huge inventories, data bases, software, and personnel who create the listings. There is lot of money and effort put into a listing. But these things become hard when the seller is a common citizen in a P2P market. The details of the product, it’s complete specifications and pricing all need to be populated manually by the seller. Add to this the work in corrections or updates or even answering questions or clarifications from potential buyers. Even big online retailers depend on past buyers to give feedback or answer questions on a product.

vendi plans to improve listings by removing human intervention. All the seller has to do is upload few pictures of the product, in this case of mobile phones.

How is this done? As covered earlier, object detection and image classification have come a long way. With state of the art AI models trained on many mobile phone brands and models, we are able to detect the phone with great accuracy. After that we deploy a series of meta data creation functions to create an accurate listing for the phone. Today, the seller does not have to do much, reducing up to 90% the work needed. The intent is with improvement in the technology built into vendi, we would be able to reduce this even further.

By continuously listing from just pictures, we have built a system that is not only removing the guess work out of the loop for sellers, but does many other tasks that help us in many other ways.

• It gives us more control on the listing for authenticity, prevent scams etc.
• Reduces user errors as it is totally automated and tested.
• The meta data creation techniques can also verify if a device was actually stolen or blacklisted etc.
• Most of all, this model dataset and the algorithms that we have built gives us a unique position and edge to deploy further assistive AI for buyers as I will cover later.

AI assistance for: BUYING

It may appear that buyers are better off in the P2P process as they see what they want, may be search quick and buy. But it’s not that easy as it looks. Just like how the seller faces the problem that listings are unique and is not part of any data base, so is the listing for a buyer, it lacks any reviews or comparisons that are assuring. So how would a buyer be able to find the right product at the right price. As I covered earlier the AI that is built for listing can also be re-purposed for visual searches, it can help in finding best matches, and with advanced techniques like scratched screen detection can easily narrow to only quality phones. Best of all this search can be done dynamically on behalf of the buyer and can span multiple sources.

A combination of visual search with text mining and natural language processing would create what we call slow search systems. We are used to search being fast and responsive, but that is when you are in front of the system and have little time. Now, how many times have the results satisfied your need, search results are only another way to filter and there is lot of work needed to create a mental model on what to buy. Imagine a system that works offline for you and helps you buy without the never ending search and comparisons. That is AI-assisted buying.

AI assistance for: SECURITY AND SCAMS

We are building an AI-assisted marketplace on vendi in stages. As stated above we already help sellers and eventually it will come to buyers and along the way these stages themselves are used to build state of the art security and scam prevention technologies. As stated earlier the greatest asset generated in these stages is the dataset. With these datasets we are able to look for patterns of scams, black listed IMEI numbers etc. We create multiple agents which assist in visual search , text mining and act as helpers. These agents in collaboration can also sniff out bad actors based on their digital foot prints. Imagine age old security tools like honeypots etc. that had helped secure networks and information, being deployed to prevent consumer marketplace scams.

To conclude, there is a whole new set of possibilities for AI assistance in the P2P marketplace transactions and with vendi we are just getting started.