With prominent platforms utilizing relevant user data to offer tailored content suggestions, could a new product with no data do it better by using advanced algorithms? And if so, how would it fit in amongst those services? As an enthusiastic startup, Cinnamon tackled these questions by looking at ways of maximizing the opportunities of today’s community-driven content creation. They quickly realized that the most logical target audience would be content creators (influencers) and their fans. With this, they built their vision. They set a deadline of just four months. But a clear gameplan was yet to be determined.
Cinnamon needed a technical partner that could substantiate every big idea, translate it into a tangible product and help the startup team incorporate processes that would assure a release. The designs were ready. What hung in the air was how to offer a sophisticated video recommendation system based on more than just text descriptions, how to translate this complexity into a comprehensive iOS and Android app experience, and how exactly machine learning (ML) would come into play.
When Cinnamon approached STRV looking for engineers to build the iOS and Android apps, the company had already acquired an ML solution for emotion recognition: AI that is capable of evaluating people’s facial reactions by analyzing the expressed emotions. Due to the immense importance of user privacy and Cinnamon’s unyielding belief that such a sensitive solution can only be utilized in specific cases—with a huge emphasis on consent, explicit opt-in and strict regulations—the team decided to omit using this solution as part of their product. However, the tool was undoubtedly one that content creators (such as Hollywood film studios that already have strict, comprehensive privacy processes in place) could potentially utilize in the future, and that was worth exploring.
To take a look at these future options, the solution was given to us for examination. Our ML team quickly recognized that the ML model’s size would require very expensive infrastructure, and efficiency was lacking. We not only pointed this out to the client but immediately came with our own solution to the problem. Cinnamon agreed, and our ML engineer was brought on board.
With ML an official part of the project, we also began digging into what it would take to build the best possible version of the video recommendation solution. Following in-depth meetings with the client, our teams realized that Cinnamon could skyrocket into something extraordinary. This led our primary ML endeavor: perfecting the product’s core service by giving users the ability to find exceptionally relevant videos at lightning speed, which is how Cinnamon was to stand out amongst the current competition.
The STRV Solution
The video recommendation engine became STRV’s biggest ML project to date. With no data available, our ML engineer had to get creative. He managed to create a dataset for the task at hand and built an AI solution that presents users with relevant videos by relying solely on a highly advanced algorithm. We then pushed the solution even further, adapting the AI to soon allow searching for videos by passing full-text queries. This resulted in two models; one processes video, the other processes text, and both create abstract representations of the content. These representations share the same properties so that items can be compared, thereby making the videos searchable without relying on the video titles, descriptions or hashtags. It is also possible to find similar videos, as they have similar representations.
With the solution ready to go, STRV took care of another critical aspect: implementation of the ML model and custom business logic to ensure that the AI performs as intended. Today, the AI runs as a serverless service, allowing Cinnamon to scale it almost arbitrarily. Once the team accumulates more relevant data about user behavior, the current business logic can be easily replaced by learnings gathered from that data.
In terms of the emotion recognition AI, we decided not to revise the solution that Cinnamon had when we came in due to saving time. Instead, our ML engineer built an entirely new embedded AI (based on recent advances in Computer Vision) that runs on clients’ devices—something that saves a lot on infrastructure costs, scales naturally and is remarkably privacy-conscious, meaning that no sensitive data whatsoever is sent over the Internet. This STRV solution ended up being 99% more efficient and 40% more accurate than the previous one. Because it can run on the client's side in a web browser or low-end mobile phone, our AI proved so valuable that Cinnamon was able to start offering it as a B2B standalone product straight away.
In just four months, STRV finalized two complex ML solutions and the full iOS and Android MVPs, helping to build a highly scalable service comparable to Instagram or TikTok in record time. Not only did our engineers, QA team and product manager go above all expectations, but Cinnamon showed phenomenal dedication to being a truly open-minded partner that wanted to work together, learn from our internal processes and give its users the absolute best. The team deserves every bit of success that’s no doubt coming their way.
“The scope of STRV’s abilities is exceptional. They know exactly what clients need and can adjust anything to adapt to those needs. The team moves companies from point A to point B.”