“In today’s landscape, avoiding AI is often not an option, its transformative potential is too significant” — Meta’s Product Manager Nisarg Shah on the Future of AI in The Recommendation Engine – AI Time Journal

As recommendation systems power more of what we read, watch, and discover, building them at scale is no longer just an engineering feat; it’s both a product and cultural challenge. In this conversation, Nisarg Shah, a product manager at Meta and a thought leader in the field, shares insights on what it takes to build billion-scale recommendation systems, how AI is evolving toward hyper-personalized and culturally aware experiences, and why the future of AI products lies in both trust and creativity.

Explore articles here: Igor Astrakhovych: From Next-Generation Drillships to a Sustainable Future for the Oceans

Nisarg, tell us how you got into building AI/ML-driven products now used by billions of people. What sparked your journey in this field?

My journey into AI/ML began with a childhood curiosity about the stock market, where I noticed patterns in price movements and realized that statistical analysis could uncover predictive insights. This interest developed during my engineering studies, where I took math and statistics courses that deepened my understanding and passion for the field. AI/ML provided the ideal platform to channel this passion, helping me build systems that solve complex problems and impact billions of people worldwide. Advancements in large language models (LLMs) and other AI techniques have expanded the possibilities in this space, enabling us to create innovative products that anticipate user needs and enhance the user experience. It’s exciting to work at the intersection of rigorous analytics and human-centric innovation, crafting solutions that transform how people connect and engage. I also can’t underestimate the profound influence my father has had on both my personal and professional life. His journey inspired me and has been a guiding force in shaping my career decisions

At Facebook, you’re leading the product for “Groups You Should Join”. From a technical and user experience standpoint, what makes this recommendation system unique?

The “Groups You Should Join” feature is key to Facebook’s mission of fostering community by connecting people with shared interests, regardless of geographical or social barriers. I lead the product development for a team of 23 engineers, 3 data scientists, and 3 engineering managers, all working toward this unified mission. We all work for one goal. What makes this recommendation system unique is its ability to build a continually updated profile of a user’s interests, blending short-term engagement signals with long-term passions. Unlike traditional systems that focus only on immediate interactions, our approach combines explicit interests, social graph, and behavioral patterns to create a multi-dimensional understanding of users. For instance, we prioritize groups where friends are active because social influence often drives meaningful connections. The system uses advanced AI to filter out noise, ensuring that recommendations are highly relevant and engaging. By refining the user profile and incorporating real-time feedback, we help users discover vibrant communities they might not have found otherwise, delivering a uniquely personalized experience.

Actually, Meta is not the first place where I’ve worked on large-scale recommendation systems. Before joining the company in 2022, I spent almost five years at Amazon, where I led the product management for Prime Video’s Watch Party. That feature became an industry benchmark, allowing people to enjoy content together virtually during the COVID-19 pandemic. I also worked on leveraging computer vision to improve content classification for Prime Video, which made discovering new content much more efficient for users. That experience really pushed me to think even bigger about scaling systems and understanding user needs on a global level. 

What are the key challenges in building scalable recommendation systems for social platforms, especially at the billion-scale? From a product perspective, what’s the hardest part?

Building recommendation systems at a billion-scale is both a technical and a product challenge. Technically, the hardest part is narrowing down millions of potential candidates to just a few hundred recommendations, quickly, efficiently, and at low cost. This requires a multi-stage ranking architecture, combining real-time processing with deep personalization. Distributed systems, advanced caching, and smart retrieval layers are essential to keep latency low while handling massive volumes of data.

Another key challenge is spotting blind spots: ensuring the system doesn’t overlook underserved cohorts or miss emerging interests. As platforms grow, so does the risk of reinforcing existing patterns and ignoring what’s new or underrepresented. 

From a product perspective, the toughest part is pushing the system into new dimensions, introducing recommendations it’s never made before. First, we leverage interests from other product features, including short-form video and community pages, and from affiliated apps to connect people with new online communities. Second, we use large language models to infer additional dimensions about a person from context, such as age group and major life events. For example, we consider signals like relocating or starting college. These approaches enable recommendations that traditional collaborative filtering or pure engagement signals would likely miss.

That could mean surfacing niche communities, latent user preferences, or new formats. These shifts require relentless experimentation, user sensitivity, and iteration to strike the right balance between novelty, trust, and long-term relevance.

I understand you also worked on launching “Interested/Not Interested” user controls at Instagram. How did you solve the problem of noisy feedback, like when a user clicks “Not Interested” but their behavior suggests otherwise?

The ‘Interested/Not Interested’ controls on Instagram empower users to shape their content experience, but interpreting feedback like ‘Not Interested’ can be challenging due to its inherent noise. For example, a user might mark a cooking reel as ‘Not Interested’ because of a particular ingredient, creator, or aesthetic, rather than a dislike of cooking content in general. To solve this, we developed a signal-processing framework that combines explicit feedback with implicit cues like dwell time, shares, or repeat views. Over time, the system aggregates these signals to build a nuanced understanding of user preferences, clarifying whether the rejection was context-specific or categorical. Industry best practices, such as ensemble modeling and iterative retraining, enhance this process by prioritizing high-confidence signals and reducing the impact of conflicting feedback. This approach ensures the recommendation system adapts dynamically, delivering content that aligns with users’ evolving tastes while minimizing irrelevant suggestions.

How do you decide which AI/ML approach is right for a given product need? What drives the choice between using LLMs, classic models, or even avoiding AI altogether?

In today’s landscape, avoiding AI is often not an option — its transformative potential is too significant. The choice between LLMs, classic models, and hybrid approaches depends on the product’s goals and the complexity of the problem. LLMs excel in tasks that require contextual understanding, like generating personalized content or interpreting nuanced feedback, due to their ability to model complex patterns at scale. Recent advancements in transformer architectures and multimodal capabilities have made LLMs even more compelling for social platforms, enabling richer, more intuitive experiences. Classic models, such as gradient-boosted trees or collaborative filtering, remain valuable for well-defined, structured tasks where speed and interpretability are essential. The decision depends on balancing performance, scalability, and user impact — evaluating whether the benefits of an LLM justify its computational cost. Ultimately, it’s about choosing the tool that maximizes user satisfaction while aligning with the product’s long-term vision.

You’ve spoken at Ai4 and other conferences. What trends do you see shaping the future of recommendation systems and AI products over the next 2–3 years?

Conversations at conferences like Ai4 highlight a shared goal: using AI to deliver unprecedented value to users. Over the next 2–3 years, recommendation systems will evolve toward highly personalized, context-aware experiences powered by multimodal LLMs that integrate text, images, and even audio to understand user intent. We’ll see systems that proactively anticipate needs, recommending actions — like joining a group or attending an event — based on real-time behavioral and social signals. A key trend is the democratization of AI creation tools, allowing non-technical users to build custom recommendation systems using no-code platforms. Additionally, advancements in federated learning and privacy-preserving AI will ensure personalization doesn’t come at the cost of user trust. These innovations will redefine how platforms connect people, making every interaction feel uniquely tailored and culturally resonant, while expanding the potential of human-AI collaboration.

On a personal level, what excites you most about working on AI products: the technology itself, user behavior, or the scale of impact?

What excites me most is the immense impact AI promises for humanity. This technology serves as a great equalizer, lowering barriers for entrepreneurs and creators globally. Imagine a future where a teenager with a bold idea can build a culturally impactful app without having to code. Or where AI-driven discoveries — potentially even a cure for cancer — emerge from systems that reason beyond human limitations. While the technology itself is thrilling, it’s the potential to empower billions — from creative kids to non-technical innovators — that truly fuels my passion. AI is set to redefine how we solve problems, create value, and connect as a society. Being part of this transformative era is both humbling and exhilarating.