EdTech: Samy Lahbabi of @Stellia On How Their Technology Will Make An Important Positive Impact On Education

Getting feedback and iterating should be done continuously in an agile mode. Except when it disrupts too much the user experience, customers value small evolution delivered very regularly much more than seldom and large updates delivering numerous evolution at once. Evolution are great opportunities to communicate with the end-user and foster change management and change.

In recent years, Big Tech has gotten a bad rep. But of course, many tech companies are doing important work making monumental positive changes to society, health, and the environment. To highlight these, we started a new interview series about “Technology Making An Important Positive Social Impact”. We are interviewing leaders of tech companies who are creating or have created a tech product that is helping to make a positive change in people’s lives or the environment. In this particular installment, we are talking to leaders of Education Technology companies, who share how their tech is helping to improve our educational system. As a part of this series, I had the pleasure of interviewing Samy Lahbabi.

After graduating in 2004 with a Master of Science in Computer Science & Electrical Eng. from the renowned École Polytechnique (Paris, France), Samy gained solid experience in innovative product management and recommendation services at a major mobile network operator and an international smartphone manufacturer. Passionate about AI and societal impact, he joined Edtech innovation in 2014, investigating very early AI to improve Education; as soon as natural language processing (the field of AI applied to language) observed a major disruption in 2018 with the advent of the Transformers Neural Network architecture, Samy created Stellia, an edtech startup with the mission to improve learning efficiency by investing deeply in AI technology specialization and designing AI edtech solutions.

Before we dive in, our readers would love to learn a bit more about you. Can you tell us a bit about your childhood backstory and how you grew up?

I was born in Morocco and grew up until the age of 9 in a highly multicultural environment. I’m very grateful to my father, who passed on to my brother and me a great passion for science and for pushing back technological boundaries. I owe it to my mother and older brother to understand that rewards and success have more to do with hard work and creating opportunities than with a simple natural gift or luck.

Although I followed in my father’s and older brother’s footsteps in scientific excellence, I’ve often been the most daring and the least inclined to run away from risk; unsurprisingly, as the youngest in my family, this was often the way to go if I was to have any chance of matching my older brother in activities.

Can you share the most interesting story that happened to you since you began your career?

In the early 2010s, I had the chance to lead some bold global projects in mobile advertising. However, although I was enthusiastic about these projects’ innovation, technology and ambition, I gradually realized that my motivation was limited by the lack of positive impact on society. So, I began to take a more selective look at the next steps in my career, preferring to swap salaries and ready-made prospects for impact and bolder bets. This choice led me to focus on startups, first as an employee, then as a founder, even though I come from a rather rational & risk-averse family. Despite the ups and downs, I’ve never regretted my career choice, focusing my efforts on social contribution and bolder bets, rather than on a safer but impact-free life.

None of us are able to achieve success without some help along the way. Is there a particular person who you are grateful towards who helped get you to where you are? Can you share a story about that?

I owe a great deal to my wife, who more than makes up for the lack of time I manage to devote to bringing up our three children and supporting me in my business ventures, especially in low times.

What’s more, at the previous startup I worked for and at Stellia, I worked with great, supportive teams, that are worth spending a lot of energy on. I can’t stress enough how mutually beneficial it is to care about others in everyday life. Not only has someone been helped, but you’ll be able to count on many of them when in need.

In 2019, I broke my leg while skiing and had to remain immobilized for almost 8 months; it’s at times like these that you appreciate having more than just work colleagues, real teammates who visit you in the hospital and then let me work from home for several months (before telecommuting became widespread during Covid).

Can you please give us your favorite “Life Lesson Quote”? Can you share how that was relevant to you in your life?

My preferred motto is “It’s those who talk the most who eat the least”.

This doesn’t mean that, like my former boss, you should always jump straight into action; you often need to take the time to analyze the situation and the possible scenarios. Nevertheless, action is obviously the only way to get results and, more importantly, feedback and lessons to improve in the future.

You are a successful business leader. Which three character traits do you think were most instrumental to your success? Can you please share a story or example for each?

I’ve sometimes been criticized for not being visionary enough, for setting out a precise vision, and sticking to it. I don’t like pure visionaries, they’re usually extremely stubborn and rigid, like the oak in “The Oak and the Reed” from La Fontaine. In the end, it is fine for the 1% of visionaries who, in most cases by chance, got it right and are celebrated; but nobody sees the other 99% of the visionaries who brought their ship to the rocks and ignored all the warnings. It’s like driving a straight line with your eyes closed. You can’t do that in real life, you have to keep your hands on the wheel and sometimes change your route when there is a traffic jam or a roadblock ahead.

I’m a fervent believer in the effectuation theory, according to which it’s certainly necessary to define your guiding principles, but in a VUCA (Volatility, Uncertainty, Complexity & Ambiguity) world, it’s above all necessary to know one’s differentiating assets and seize opportunities that are often unexpected.

In the early months of our startup Stellia, we wanted to build AI solutions for Education, but we lacked sufficient in-house experience, time, and a great idea. So, we simply started by providing AI training. Only then did we realize that our AI teachers didn’t have enough time to answer all the students’ questions, and they were tired of answering the same questions over and over again. Meanwhile, by continually enriching our AI courses with practical workshops, our team has strengthened its expertise and agility in AI frameworks and models and the understanding of their potential. We realized the great need for a tool to answer students’ most basic questions; with greater expertise to crack it, the time had finally come to shift our efforts from AI training towards our company’s mission and start building a chatbot for students!

A second of my traits is to evaluate each piece of information in terms of likelihood, rather than as true or false, i.e. a binary or Manichean vision. Everyone prefers clarity rather than ambiguity, but unfortunately, the world is complex, and you should use fuzzy logic rather than a computer, approximating everything to 0 and 1. You should add grains of salt to each piece of information, adjusting the volume of salt according to the credibility of the information and its sources, of course avoiding becoming a conspirator either! The downside is that this scientific approach can block or delay your action as you never fully trust any information; at least when your probability estimation is not off, you know what the risk of action is, and can therefore decide better while being better prepared for the risk that the information may be wrong. You need to counterbalance this trait with a riskier attitude to avoid getting stuck too often, but you’re always more reactive in the face of risk. I’ve worked with many people ready to decide instantly, once and for all, whether each new piece of information was true or false. In many cases, they made the wrong decisions and stuck to them for a long time, with serious consequences, before finally agreeing to change their minds.

This trait is bound to become increasingly important in the face of fake news and multipolarity in a more and more VUCA world.

I also really appreciate benevolence, which is essential for teamwork, mutual support, and the generation of opportunities: it’s by genuinely interacting with people that you’ll sign contracts or generate fruitful partnerships!

Let’s now shift to the main part of our discussion about the tech tools that you are helping to create that can make a positive social impact on our educational systems. To begin, what problems are you aiming to solve?

In our societies, the 3 main high-value asset types are:

1. human-being/human resources/workforce/manpower,
2. knowledge/information
3. and physical assets/means of production

A company’s value is primarily determined by the quality, quantity, and potential of its proprietary assets in these three areas. The company’s potential comes essentially from the interactions between these assets. However, this potential is severely limited by friction due to the very different nature of these assets. Unfortunately, as there is often no equivalent to Google/ChatGPT for reliable, contextual knowledge of a company or training course, we waste a lot of time finding relevant in-house information; then, many falls back on the Internet and thus unreliable, non-expert information.

Stellia’s mission is to facilitate and stimulate interactions between human beings (1.) and the qualitative knowledge from organizations (2.), which is also essential to enable human beings (1.) to better exploit physical assets (3.). For that, we conceive AI knowledge chatbots to identify, contextualize, and push the most relevant knowledge to each student or employee according to their needs and expertise. To achieve this goal, we develop both our solutions and, thanks to our strong AI team, our own AI algorithms and models, specializing and optimizing the best open-source AI models for our tasks.

To reach the highest level of reliability and relevancy, we have chosen to conceive our solutions on the strong promise to deliver answers exclusively based on bodies of knowledge from or certified by our clients.

Working with and for education institutions, training organizations, and corporations, we rely on our customers’ or partners’ expertise in their field to define the scope of qualitative knowledge.

Our tailor-made AI models extract, process, and model knowledge from our customers’ documents and videos, then propose only verified knowledge that matches the pedagogical context of the training course in progress or the company, and the needs of the end-user. This allows us to adapt the level of complexity to users’ skills, based on the knowledge extracted directly from content associated with their context. This approach avoids reliance on external knowledge that does not come from or has not been validated by the customer beforehand. Consequently, we don’t answer questions that go beyond the input knowledge corpus, rather than proposing a risky answer, as other generative AI tools might.

Our second axis is to deliver rich answers to stimulate curiosity and critical thinking. So, in addition to a synthetic and generative answer, we highlight all the sources on which this answer is exclusively based, with extracts of documents offering more context.

After identifying and linking the main concepts in the source corpus, we generate a knowledge graph, and pass on to the end-user the main concepts related to his question, to help them find their way around the knowledge. Instead of settling for a concise answer, users are encouraged to go further in exploring and enriching their knowledge.

How do you think your technology can address this?

Adaptive learning, which involves adjusting each learner’s learning path in real-time to enable them to make the best possible progress based on their current understanding, has long been an unattainable dream for Education. However, deploying adaptive learning means analyzing what each learner’s interaction with the content means for their understanding of each concept. Adaptive learning can only be deployed on a large scale if you can automate this analysis, which should be based primarily on the relevant metadata attached to each learning resource.

In my previous venture at a startup, I was responsible for exploring how AI could automate metadata generation and to be honest, the results weren’t consistent and convincing across the knowledge domains.

End of 2017, Google Research labs published a paper “Attention is all you need” about a novel AI Deep Neural Network architecture called Transformers, which was far more efficient in language processing; a few months later, some implementations of this architecture, such as Bert (Google) or GPT-3 (OpenAI) demonstrated performance on-par or even above human performance on some tasks, such as closed-domain short answering. This event triggered the creation of Stellia to focus on this disruptive advance and specialize it for Education and, more broadly, expert knowledge.

Our first AI knowledge chatbot was released mid-2020, matching student questions with Q&As from a FAQ. We then started exploring generative AI, but here again, despite our important efforts the performance was not yet sufficient to match the stringent requirements of an extremely high relevancy for Education. Boosted by RLHF (Reinforcement Learning with Human Feedback), the release of GPT3.5 in November 22 (which powered ChatGPT that was made available to the general public) finally brought generative AI on natural language to high performance and the limelight.

However, Large Language Models (LLM) such as GPT3.5/4/4o, Gemini, or Mistral alone do not solve the problems of knowledge access and deep contextualization. Since 2019, we’ve built a rich AI workflow to better extract knowledge from documents and videos, chunk and model it, and, given a question, retrieve the most relevant knowledge units. We use LLM mainly as the last mile to synthesize an answer and better contextualize the knowledge to the user questions. This approach, called RAG (Retrieval Augmented Generative AI), enables focusing LLM on specific pre-identified knowledge, instead of relying on generic but unreliable knowledge from massive public sources. In addition to knowledge graph modeling and, more recently, the fine-tuning of our LLM, it is this RAG approach we have been betting on with since our creation; RAG approach has been progressively gaining momentum in 2023, being now recognized as the most relevant one to boost relevancy and avoid most LLM hallucinations.

Even though our solutions have already achieved high relevancy with a limited number of errors compared to most competitors, we continue to work on further improvements around the performance and versatility of our AI models and workflow; we strive to ensure that we always respond better to user questions based solely on the knowledge that our customers trust.

Our vision is to enable our solutions to build and interact with users in an ever more personalized way, taking more context into account, better understanding user needs, and proactively offering personal recommendations. This objective will be pursued while preserving the confidentiality of personal data with partners who are experts in this field (state-of-the-art differential confidentiality and edge computing), such as Cozy Cloud.

Stellia is particularly well positioned to realize this vision, given its highly relevant approach, its technological mastery, its great partners, and the trust placed in it by major institutions (the French Ministry of Education, the Banque de France, Arizona State University, the University of Illinois…) in the face of dominant players who inspire great fear and mistrust (Google, Microsoft, OpenAI…).

The only way to deploy Generative AI on a large scale is to become more professional and drastically reduce errors, which our RAG approach solves, and to become more personal to provide highly contextual answers to each user. After Gen AI, Personal AI is the next AI revolution!

Can you tell us the backstory about what inspired you to originally feel passionate about education?

You can remain focused on yourself and work only for your own benefit; however, looking at the many alarming challenges facing our world and societies, if you have the opportunity, you can’t remain passive, and we should try to mobilize everyone’s energy. As Education is aimed at people from the very beginning of their lives, Education is of particular importance in making a major contribution to the behavior of every citizen in our world. I am truly convinced that better Education of people contributes greatly to more responsible behavior and to mobilizing people in the right direction to meet all these challenges.

Internet and Wikipedia have opened the door to free knowledge to most of humanity, but unfortunately, as a decentralized/uncontrolled service, there is a huge problem of reliability and much knowledge remains private.

MooCs (Massive Open Online Courses) have paved the way for free education for all, but learners still struggle because MooCs’ freemium business models imply limited support and personalization.

Today, at Stellia, we see a huge opportunity for AI solutions to cross the last mile, to personalize and support everyone in their learning journey, in contexts without or with teachers or trainers.

At this stage, and for years to come, there is a huge need for teachers and trainers to support learners in their most complex demands, to understand them better, especially taking into account a broader qualitative context, and to animate teaching in a livelier way. I don’t believe that, at least within the next 10 years, Artificial General Intelligence (AGI) will be available, i.e. that a single AI will outperform humans in most reasoning tasks. I don’t trust either anyone who bets on when it might be available.

But I trust our startup Stellia in becoming the personal knowledge coach of every learner and employee!

How do you think your technology might change the world?

Changing the world means allowing everyone fair access to reliable education, culture, and quality information. This is the only path to a better society and addressing the challenges humankind is facing, such as climate change, inequalities, and mistrust in governing institutions, …

Technology, mainly AI, is a great tool to pursue these goals more effectively. That’s exactly what we do at Stellia with our Knowledge Assistant, which provides reliable and contextual knowledge, tailored to each learner and employee.

Keeping the “Law of Unintended Consequences” in mind, can you see any potential drawbacks about this technology that people should think more deeply about?

Clearly, the development of generative AI will have consequences that are not yet fully known today, but whose contours we are beginning to see. In particular, we’ll have to be increasingly vigilant about the quality of information, whether textual, video, or other. The internet will be massively diluted by generated content, some of it extremely misleading and manipulated for highly malicious purposes, much of it simply worthless content, simply aimed at promoting businesses by fooling search engines and LLMs.

While there is little doubt that citizens will not be overtaken by AI in the short to medium term, we will need to further develop our critical thinking skills to evaluate information. Beyond education and learning, there is no better and more scalable solution than harnessing the tools of AI to meet this challenge.

What’s more, we must also take AI biases into consideration, and in a preventive way, continue to avoid as far as possible the unforeseen consequences of our solutions upstream of their deployment, but also remain extremely reactive to any alerts on this subject.

Stellia is engaging all its capabilities in this direction.

How do you envision the landscape of education evolving over the next decade, and how does your technology fit into that future?

First of all, provided it is correctly implemented, AI can be a unique opportunity to offer everyone the initial support they need to learn and progress, based on reliable knowledge. Khan Academy’s Khanmigo or Stellia knowledge assistants, deployed for instance at Arizona State University in Math or at the University of Illinois on video courses for Business, are excellent examples of efforts in this direction.

The next step is Personal AI to further personalize support, from struggling students to unmotivated employees to the most curious or demanding. Everyone is different!

Tomorrow, everyone will have a secure AI knowledge coach, educated to the specificities of each individual, drawing on reliable and expert knowledge, while keeping all data extremely secure and personal data private.

This is the next stage in the qualitative democratization of Education at scale!

Based on your experience and success, can you please share “Five things you need to know to successfully create technology that can make a positive social impact”?

1. Your greatest asset is your team. Make sure you build a diverse, complementary and smart team, capable of playing as a team! And make sure you hire people whose values match those of the company, especially when it comes to impact. At Stellia, we have hired a full-time human resources manager to help us recruit better, ensure that all team members give their best collectively, and that everyone is happy and motivated in their work.

2. Clearly, you need to start with the end-user, finding out what their needs are, or what recurring frictions they have. Starting with an idea and analyzing the competition can be a bad idea. This should be a next step, otherwise, your mind will be skewed by a mixture of existing solutions, which may not solve the problem to your satisfaction and not allow you to propose a differentiated solution. If you can see yourself as a typical end-user, all the better but be sure to check that your needs are also shared by others.

In Edtech, for K12, you need to have young kids and ask them if you are targeting their class range! Your interns and young graduates could also be a better source than you for higher ed!

Last point, ensure your user questionnaires are neutral and not unwittingly suggest to the interviewee one solution instead of keeping it open.

3. Your key assets for a great product are:

1. obviously, great features!
2. a great frictionless User Experience. But because it’s visible and therefore replicable, once you’ve reached a certain notoriety, you’d better be highly viral to maximize your market share before you’re copied
3. developing great technology. Costly, and hard to architecture something very differentiated
4. having a great ecosystem of partners. Less replicable, but very long to build, costly in time and man.power and hard to convince partners unless you are already successful. So this is rather naturally self-reinforcing only once you already get some traction

At Stellia, we only recently double-downed on UX (b.) which we hadn’t time to sufficiently take care of. We invested mostly in technology (c.), but you have to find the right differentiated positioning when other BigTech companies are spending hundreds of millions of $!

For partners, there are plenty of potential partners (d.) with which you could build a better value proposition. There are even too many, and you might be frustrated being forced to focus only on the few one that matter the most for you!

At Stellia, it was a tough choice! Beyond a cost/benefit analysis, it is also a matter of human fit, alignment with your roadmap, and the respective motivation.

4. If you are serious about impact, there is no question you should try to measure it! God knows why, precisely measuring impact is always far harder to set up than simply counting usage. It can also be very intrusive if you are willing to collect the user’s full context before, during, and after the considered experience. So, you need to find there some compromises: limited measures on a large scale and in-depth data collection from a few representative users, optionally thanks to surveys, from which you can ideally extrapolate to the whole end-user population. At Stellia, we are selling our solution to training organizations so we do not have any direct interactions with the end-users.

Be sure to focus on metrics that accurately reflect a positive societal impact. For example, in France, the government was willing in the last century to ensure the largest share of the K12 students would reach the baccalaureate level. In the end, it led to the teachers lowering the baccalaureate level and greatly increasing the student failure rate at the university level. For our student chatbot, measuring the number of questions asked by each student does not reflect the volume of questions that were not pushed to teachers, as students might ask a chatbot many questions they would not have asked their teachers, and maybe vice versa.

5. Getting feedback and iterating should be done continuously in an agile mode. Except when it disrupts too much the user experience, customers value small evolution delivered very regularly much more than seldom and large updates delivering numerous evolution at once. Evolution are great opportunities to communicate with the end-user and foster change management and change.

In the realm of EdTech, there’s often data collection involved. How do you ensure the ethical handling of user data, especially when it concerns students?

First of all, Stellia is highly aligned with the EU GDPR and EU AI Act as guiding principles. In particular, it encompasses explicit user consent, data minimization, documentation of data processing, user rights over personal data, data security, and vulnerability information… which is the minimum we can expect from a citizen’s point of view.

As far as AI is concerned, there is always a dilemma between minimizing data collection and collecting everything possible to best feed the models.

We’ve also all heard of AI bias, which theoretically seems easy to solve if you can ensure that your input data (usually created by humans) is completely neutral or not biased in any direction and that the diversity of your training data is representative of your user data. However, this reasoning overlooks borderline cases, which are rare among your users and in your training data if they are representative. You can try to over-represent them in the training data sets, but then you introduce a bias in the other direction, such as what Gemini image generation did. The only solution is to have a representative data set that is extremely large enough to still contain enough data for minorities or marginal cases. This partly explains the great performance of Large Language Models.

At Stellia, we believe that we should use personal data carefully, only after having explained to the user how we use it and after having obtained his explicit consent. But personal data can enable us to better understand a user’s context and needs, and thus serve them better with contextualized knowledge aligned with their situation.

If you could tell other young people one thing about why they should consider making a positive impact on our environment or society, like you, what would you tell them?

Sorry, I’d rather say 2 things:

1. Start putting most of your energy into generating a positive impact on society and others as soon as you can!
2. Keep learning throughout your life to continuously grow your mind, instead of wasting it on silly activities such as video games that are repetitive or browsing social networks!

Is there a person in the world with whom you would like to have a private breakfast or lunch, and why?

Previously I would have said Elon Musk, but, wow, how disappointing his behavior has been in recent years!

Now, I would hesitate between Salman Khan or Luis von Ahn, as my kids would love to ask him for a free Super Duolingo subscription, but I am not yet willing to offer them!

I would also be highly delighted to share some time with old friends who are now very far away.

How can our readers further follow your work online?

Keep an eye on https://www.linkedin.com/company/stellia-ai/ !

Thank you for a meaningful conversation. We wish you continued success with your mission.