Refsee: from Automating Monotonous Work to Helping Students
Refsee: from Automating Monotonous Work to Helping Students

Refsee: from Automating Monotonous Work to Helping Students

18 january, 20236 minutes to read
Follow [Durov's // Code] on Telegram

Process automation not only simplifies work, but also saves time, which we can later invest in creating new products and projects. Peter Bondarenko and Mitia Muravyov kept that in mind when creating a reference search service - Refsee.

Durov's Code talked with creators of the project about its they key stages and future development.

— Could you please tell us what were you doing before you started the project and how the idea came about?

Peter: Mitia and I have been running fancyshot.com - a creative agency and production - for almost 10 years, and we are still doing this. During this time, we have worked with almost all major clients in Russia. Since there were not so many employees in the company for a long time, Mitia and I took care of most of the issues related to the business and therefore gained expertise in many areas. In fact, we have always been actively involved in creating new things, so we have come up with many ideas and made thousands of presentations.

The history of Refsee began when Mitya and I were once again preparing a creative concept for a client. At that moment I noticed that the process of finding an idea is an incredible pleasure, but finding suitable references and creating a high-quality presentation eventually turns into an unpredictable routine. Previously, we had a bunch of bookmarks and likes on Vimeo and similar sites, where we crawled through the pages from project to project to find the right reference or idea.

I thought: “Why not automate this?”

I immediately thought about neural networks, since I had been fond of them for a long time, and in a few months I created the first prototype that could recognize objects in a video. Then, when it became clear that it actually worked, I made Telegram bot so that I could quickly test the future product. I showed it to several directors and cameramen, and the reaction was unambiguous - we need to make a full-fledged product. The thing really works and is extremely useful.

Then I made the first beta version of the site using Python + Django, and then uploaded a story to Instagram with the announcement. As a result, we had about 600-700 during the first month, and I began to collect feedback from them. Since these were users who came from my personal Instagram account, I knew them personally and asked each of them some  questions about how to improve the product and what was done poorly at that time.

— Refsee as a product, how it appeared and who created it in the form that it has now.

At some point, Alexey Shaikhaleyev  joined the project and took responsibility for the servers, the technical part, scaling, and so on. In general, he ran the technical part of the product.

From that moment to this day, I am adding more and more new features, based on user feedback. Mitia builds relationships with clients and funds, manages marketing and recruits people to the team. And Lesha makes sure that all this works stably, quickly and pleasantly.

— It is very interesting to understand how your neural network works. How much data you have in the database and how fast does the search engine work? Do I understand correctly that the search engine works on the server - and not on the device?

Peter: In fact, the service consists of two large parts, and each of them has its own set of neural networks.

The first part is a bot that checks weekly for new videos in a certain set of sources. As soon as it finds novelties, the video gets into the processing part, where it is divided into scenes and, using several neural networks, the best frame is selected in order to add it to the database. The same bot converts images into a special data format understandable to the search engine.

The second part is the refsee.com website itself. There is another neural network here, which is attached to the search bar. The neural network allows you to convert user requests in any language into such a database query in order to get the most relevant response.

— Now you say you have more than 10,000 active users per month. And among your clients are famous brands. How did you come to them (or rather, they came to you?) Were you able to get good feedback from them and improve the product?

Mitia: The first major brands and clients came to the site simply as we continued and continue to work with the agency and, whatever one may say, we are just have connections in this sphere. In the same way, agencies from all other the world began to come to us. At first, these were just company employees, and then we were approached by design makers from companies who wanted to make a purchase. Now, there are more than 1000 advertising agencies and productions from all over the world in our client pool.

Recently, we became partners with the largest German film and advertising university - Filmakademie Baden-Württemberg. Warner Brothers also bought a yearly corporate account from us.

We regularly call up such clients and conduct cust-dev interviews. Indeed, such conversations give us a lot of useful information about product improvement and even business organization. Users, in general, are very fond of the service and are always very willing to share their experience and sincerely want the product to develop and grow, which is, of course, super nice.

— At the core of your product is a web interface. Have you thought about creating an app for smartphones and tablets, maybe even making native apps for computers?

Peter: Yes, they thought, and even received several proposals to make an application for the MacOS . But somehow these tasks did not become a priority for us and we have not yet begun the creation process. Rather, as more and more people join the team, we will definitely make both a mobile and a native application for computers. Now we are focused on the current product and expand user flow.

— You have a rather small team. Only three people. How do you share responsibilities and have you thought about hiring more assistants to scale?

Mitia: That's not entirely true. We have three directions and, accordingly, three people lead them. Frontend, translators, copywriters, and even performance is done under our control, but not by the three of us. We are also very lucky, we have some of the strongest advisors from international giant companies in C/C-1 positions. Their experience and advice is really unique, it helps us move as efficiently as possible. Nevertheless, now we are actively working on expanding the team and scaling the business.

Peter: At the moment, the secret of success was that the product model works with the help of bots that do all the routine work. I peeped such a scheme from Pieter Levels - the author of the book "MAKE: Bootstrapper's Handbook"

— Refsee in 5 years, how do you see it? What are the plans for the development of the project now?

Peter and Mitia: In 5 years, Refsee will be a full-fledged international company. Definitely, I would not want to be tied to one point in the world, because the product itself does not require this. Therefore, I am sure that the team will be from all over the world.

The product will develop towards a kind of universal tool for the creative industry. Such a kind of a Swiss knife that will help in creating content from the moment of the very first idea to the final presentation to the client. We are already working on adding tools for creating storyboards, moodboards and presentations.

Moreover, throughout the history of Refsee, an incredibly interesting and powerful audience of creators, operators, directors, productions and agencies has gathered around the product, so it's time to add elements of a social network with public profiles, likes and portfolios to the product. This will allow sharing experience between creatives and moving the industry in general.

18 january, 2023
Follow [Durov's // Code] on Telegram