Nir Bar-Lev is the CEO & Co-Founder of Allegro AI. Allegro AI specializes in helping companies develop, deploy and manage machine & deep learning solutions. With Allegro AI, organizations bring to market and manage higher quality products, faster and more cost effectively. The products are based on the Allegro Trains open source ML & DL experiment manager and ML-Ops package.
What initially attracted you to AI?
What I have been mostly drawn to in my career has been about bringing cutting edge tech innovation to address problems or opportunities (and actually they are two sides of the same coin) on a huge scale. I must admit that my time at Google has certainly helped shape this inclination.
AI certainly ticks off both these boxes. It is at the cutting edge of some of the technology frontiers today and it has the potential to affect almost every single aspect of our lives on this planet.
You have had an impressive career starting at Google as the founding product lead for Google’s voice recognition platform. Could you discuss these early days of working at Google and what you learned from this experience?
Coming straight out of business school from the Wharton School of Business I was struck by how Google was functioning at extreme odds with established business norms on how to run successful businesses, as taught in the best business schools in the world and as I experienced in my career pre-business school. I vividly remember discussing this with a couple of my colleagues who also joined Google at the same time straight out of an MBA.
It turns out Google changed – to some extent – the business playbook, but it also enjoyed an immense virtual firehouse of money from its ad business that allowed it to experiment in ways that most companies could not afford to do. I can attest that as I spent a decade at Google it increasingly adopted “mainstream” established business practices and thought processes as it grew.
To me also, leading the voice recognition platform as the head product manager, I had to work with research scientists. This was actually one of the earliest, if not the first, research team at Google that was really about applied research. To me this was a big challenge. Researchers have very different mindsets than engineers and here I was trying to work with accomplished researchers in a company that is extremely engineering oriented.
Turns out that the challenges I faced back then almost 15yrs ago are very similar to problems companies face today when trying to assimilate AI data scientists into their organizations.
In 2016 you proceeded to become a Co-Founder of Allegro AI? What was your inspiration behind launching Allegro AI?
In founding Allegro AI, I teamed up with two amazing partners who are out of this world engineering talents. One of my partners was the first PhD student in one of Israel’s first and currently leading AI labs in what is arguably one of the leading AI hubs globally. So he really – to me – was part of the founding teams of applied AI in the community locally. He had the vision to see how applying ML / DL in practice would have to deal with a new set of challenges around scale, automation, reliability, quality and more. In talking to them it became clear to me that I can contribute to the team from my experience at Google and earlier to really have a shot at creating a company that can have an immense impact on AI through the tools we provide. Google and some of the other tech giants are in an enviable position in terms of their ability to garner endless resources of the best quality at these challenges. But pretty much everyone else cannot afford that (whether in terms of access to talent, monetary resources, company focus, etc). So this was an opportunity to aligned exactly with what I love to do most (see q1) and help the whole ecosystem.
Allegro AI serves as an open source machine learning & deep learning management platform. Could you discuss the benefits of using open source software?
Open source has several benefits to it. Most importantly it leverages the wider community to improve the product itself. Users find bugs, issues, there is a wide discourse on features that are of interest; integration into other [open source] tools is much easier to facilitate than it would be b/w two commercial organizations with closed source proprietary tools; etc.
It provides a great model for a win-win for both the community and the company behind it. It lends easily to trying and testing and even expanding for organizations that do not / will not pay, and at the same time enables larger potential customers to pay for extended features / services based on top of a widely used (and therefore less risky) piece of software.
Allegro AI offers data management services. Could you discuss the type of tools that are offered for this?
Allegro Ai offers both structured data and unstructured data management. However, whereas there are a host of proven structured data management solutions, we provide a unique solution to unstructured data.
Specifically, it is important to qualify the type of data management we provide. The idea is not physical data management but rather data management from an AI angle. For AI, it is critical for the data science team to understand what data they have at their disposal. With unstructured data that is quite difficult. Imagine thousands or hundreds of thousands of hours of video, or audio. Imagine billions of sensor signals, etc.
Data scientists need to know the variance of their data to align with the different situations so they can effectively train their models. They need to understand if there’s critical pieces of data that are missing; if there are biases or skews in the data.
And then – on the flip side – they need to have tools to address these situations cost effectively and quickly without having to go out and source new physical data and annotate / label it (a very costly and time consuming undertaking).
This is in essence the type of tooling we provide around this area: powerful tools to do “AI BI (business intelligence)” on your data at an unprecedented level of granularity and detail and on the flip side tools to tightly integrate the data into the experiments and models such that with zero code data scientists can set up effective training runs with the data at hand.
On top of that we provide additional value-add in optimization of data flow, data move etc. Since we are talking about processing terabytes of data. Moving it around is expensive and companies need a solution to optimize that as well.
Allegro AI also offers the outsourcing of data engineering services. What are some of the offerings that are available?
Allegro Ai is primarily a product company and we see ourselves providing the tools, infrastructure or scaffolding for companies to develop, depley and/ro manage products with Ai (DL / ML) models integrated in them.
That said, this is a new area and our customers at times need help setting up their specific pipelines built on top of our tools, or even help with jump-starting their models themselves. When these situations happen, we provide ancillary services to our core s/w offering.
Could you discuss the importance of Federated Learning and how Allegro AI can be used in this context?
Federated learning is basically the ability to train a single AI model leveraging (trained on) datasets located in different physical locations without bringing those datasets to a single location. We also provide an enhanced version of that, which we call “blind federated learning” or “blind collaborative learning” where no single entity in this scenario has access to data that does not belong to it, including the entity that gets the ultimate model.
Federate learning is important in various situations where data privacy or regulatory or IP / confidentiality is critical to preserve while at the same time there is interest to leverage different datasets. For example, two or more hospitals or medical institutions that want to collaborate on training a model for CT scans; or two governmental agencies that want to collaborate on homeland security data to build some anti-terrorism model but for legal reasons cannot expose the data even to one another.
Or even situations where a single entity cannot move its various stores of data b/c it is prohibitively expensive – for example a global automotive OEM looking to train autonomous vehicles leveraging data collected from cars driving all over the world.
Allegro AI is one of less than a handful of companies world-wide that has a proven and tested commercial platform that facilitates federated learning.
Is there anything else that you would like to share about Allegro AI?
Allegro AI is a rising force in the world of AI tools and ML-Ops. Just this past quarter, during the midst of the first wave of the covid-19 crisis we experienced growth that more than doubled our customer base in just that 3mn period.
Thank you for the interview, readers who wish to learn more should visit Allegro AI.