Provectus CTO on how enterprises can boost AI implementation

Provectus is an IT systems integration and consulting firm that specializes in one thing only: AI. Most recently, the company announced a partnership with Tecton, which provides a feature store for building AI models that’s based on open source Feast software.

Under terms of that alliance, Provectus will focus on making it simpler to deploy Feast on the Amazon Web Services (AWS) cloud, in addition to helping define common, non-opinionated vendor-agnostic application programming interfaces (APIs) for feature stores.

VentureBeat caught up with Provectus CTO Stepan Pushkarev to get a better sense of what it takes for organizations investing in AI to really succeed.

This interview has been edited for brevity and clarity.

VentureBeat: What’s you’re best AI advice to organizations right now?

Stepan Pushkarev: Do more machine learning models. If you just choose one idea and try a proof of concept for the next year, you will never reach the goal. Choose to prioritize 3-5 use cases and just go for it. Carve it out into a separate cross-functional team that is able to execute fully all on its own. From a technology perspective, combine all the best from cloud vendors and open source. You don’t need to buy the platform.

VentureBeat: What makes an AI project any different from any other IT project?

Pushkarev: Obviously, AI is software at the end of the day. But it’s not fair to say that AI is the same software; it’s a new type of software. It has its own flavors. Companies that tend to do traditional IT projects very often struggle with the implementation of AI projects. It really depends on the maturity of the company and the level of innovation and experimentation culture of the company.

VentureBeat: One of the cultural issues that organizations are struggling with is that the pace at which a data science team works doesn’t always align with the faster rate at which applications are being developed. How do we bridge the divide between these teams?

Pushkarev: It’s a multi-diamond dimensional question. There is no straightforward answer. There are many cultural aspects. To ship machine learning projects faster, you need the appropriate infrastructure so that you can run experiments faster. You need to track those experiments as they roll out to production. This is something that is being developed at this moment. Companies do not often have the right machine learning infrastructure in place to perform at the pace they want. How data science projects are being managed is always a question. There might be many cases, for instance, where the data scientists are being put in a separate silo and they just operate on their own. Obviously, those projects will never see production. We usually recommend putting data machine learning engineers closer to the business to basically embed them into the feature teams or product teams so they can have the full context. They work on the same sprints on the same iterations to ship software to production.

VentureBeat: It feels like a lot of AI projects for better or worse have been accelerated since the start of the economic downturn brought on by the COVID-19 pandemic. Are organizations really up to that challenge?

Pushkarev: Everybody focused on productivity because there is no chance to wait anymore. I believe it also depends on the maturity of the company. The most important thing is to have a specific business need and use case with a clear ROI (return on investment). That’s probably the main driving factor for the projects going to production. If you don’t have that strong business use case, it’s just going to remain a proof of concept in many enterprises. You will not have the same quality control for those projects. That the first-class problem in enterprises.

VentureBeat: Do businesses have unrealistic expectations of AI?

Pushkarev: We definitely have to set the expectations of the business executives about AI. For the mature companies with an established innovation culture, they know what they want and know what they do. The only thing that they need is just the right tooling. However, for other enterprises, there is a need to take two steps back to work on strategy and make adjustments to their organizational structure. Leaders define goals and prioritize business use cases and only then start executing pilots. Education of executives focuses on what ML (machine learning) is, what you could expect from the machine learning model. What sort of accuracy to expect? What is the average adoption cycle for the machine learning solution in the enterprise? That’s what we call a management boot camp for AI. During this management boot camp, we talk about things like budgeting and total cost of ownership. The cost structure for machine learning solutions differs from traditional IT projects because of the cost of high-quality data, the cost of training, retraining, the cost of machine learning inferencing, the cost of new types of specialists, such as machine learning engineers. Estimating machine learning algorithm complexity in the early stages is also crucial because business might think about something that is in academia that is not yet real.

VentureBeat: Are best practices for machine learning operations emerging?

Pushkarev: That’s one of the hottest topics these days. People really understand this idea quite well. MLOps is usually just basically the pipelines between experimentation and development and production.

VentureBeat: Do you think MLOps will remain a separate discipline or just become part of IT operations?

Pushkarev: It will eventually converge. But at this moment, I’m not sure about the timeline. There are still lots of gaps.

VentureBeat: What mistakes do organizations most commonly make?

Pushkarev: People get stuck in all those PowerPoint presentations. That lack of action is something that prevents companies [from moving] forward quickly. Bias to action is very important. Companies just spend weeks organizing the meetings.

VentureBeat: There’s a perception that these projects require a global systems integrator. What makes smaller integrators such as Provectus more attractive to work with?

Pushkarev: First of all is our focus. We are good at one thing that we do really well. The ability to execute and have a low-level type of discussion alongside macro-level business discussions is something we’re really good at. We are recognized by our partners, such as AWS, and unlike other professional services companies, we’re also an active contributor to open source, machine learning, and data infrastructure projects. Our clients range from cutting-edge startups to midsize and large enterprises seeking innovation through AI. We work on the hardest problems in the world, [everything from] simulating turbulence in supernova explosion to demand forecasting. We’re equally deep in technology, business, and strategy. We can start from business objectives and deliver a working AI solution in a matter of weeks.