Personalizing food preferences is pivotal to grocers operating at the convergence of data and human emotions. The core of a grocer's operations revolves around food, and maximizing customer satisfaction by catering to their taste preferences is crucial for retaining their loyalty. Extensive research highlights the pressing need to address the profitability challenges faced by grocers in the online domain and grocers must take proactive measures to enhance their online platforms.

Taste Intelligence platform Halla harnesses the power of artificial intelligence (AI) and machine learning (ML) to deliver real-time insights with advanced technology and domain-specific knowledge. Halla leverages AI-driven technology to empower grocery retailers in making data-backed decisions that drive online profitability. With solutions ranging from personalized recommendations to boost incremental sales, enhanced search functionality for an optimal user experience, and efficient order substitutions, Halla enables retailers to meet customer demands seamlessly. Moreover, Halla enriches the shopping experience by providing shoppers with personalized grocery curation recommendations.

In this exclusive interview, Grocery Doppio and Halla co-founders Gabriel Nipote (COO), Henry Michaelson (CTO), and Spencer Price (CEO) discuss the company's vision and objectives. They shed light on how their innovative approach revolutionizes the retail grocery industry by offering comprehensive insights through customer data interpretation. As a result, grocers experience a substantial boost in incremental revenue, sometimes reaching up to tenfold.

Grocery Doppio: Tell us about the company vision and what led you to find the company.

Nipote: To provide context and background, we are childhood friends from Los Angeles. We attended the same high school and were known for our ambition, entrepreneurial spirit, and interest in technology and mathematics. Henry and I shared math classes, while Spencer and I formed a small group that convinced our IT director to teach us Java during lunch breaks because our school didn't offer computer science courses at the time. After high school, we went our separate ways, Henry at UC Berkeley, Spencer at USC, and I went to UCLA. However, we reconnected during college and decided to start a company together.

The kernel for our venture stemmed from our shared belief in the future potential of AI and data to drive robust recommendation systems. While machine learning was a trending topic, we noticed an under-serviced market in the food industry. We recognized a need for better understanding and catering to people's preferences through large-scale recommender systems. Initially, our primary approach was to create a consumer-facing app that resembled platforms like Spotify or Yelp. However, our app had a distinct focus on providing personalized restaurant dishes recommendations. We raised angel funding in the early stages; however, shortly after, we recognized a significant opportunity in the online grocery sector. We realized that personalized recommendations could be applied to various use cases of the online grocery experience, including product suggestions, search substitutions, and retail media. Our goal was to optimize the shopping experience by tailoring product recommendations to each customer's taste preferences and buying patterns. We have a strong passion for data, computer science, and AI; we pivoted our company to align with this focus.

Michaelson: Our journey has been evolving, and our central focus has been at the intersection of AI and food. While the specific applications have evolved, we have accumulated a wealth of knowledge and expertise in developing machine learning algorithms and understanding the crucial features necessary for effectively deploying AI in the food domain. Our work entails leveraging sequence-based user behavior to enhance recommendation systems. Additionally, we place great emphasis on domain knowledge, utilizing it to glean powerful insights from data and further enhance the capabilities of our AI solutions. 

Grocery Doppio: Can you explain the architecture of your personalization system?

Michaelson:  Our underlying architecture at Halla is designed to be a highly scalable real-time AI system. The system efficiently processes large volumes of raw data, allowing for long-term analysis at regular intervals. We ensure data efficiency by normalizing the data format and employing fast lookup indices for real-time data storage, enabling quick retrieval and decision-making.

In practical terms, this means that when users interact with the platform, such as adding items to their cart or conducting searches, the system can rapidly adapt and provide relevant recommendations. For example, if a user recently viewed tomatoes, their search results and suggested substitutions would dynamically reflect this information in real time. Our objective is to process and utilize data swiftly to facilitate dynamic decision-making.

Our approach is similar to how Spotify builds its recommendation engine. Just as Spotify enhances people's music experience with AI, we aim to delight shoppers by leveraging AI and their interactions to deliver seamless, personalized, relevant recommendations in the grocery industry.

Grocery Doppio: How do you ensure the accuracy and reliability of your system's recommendations?

Michaelson: We follow a systematic approach to monitor and evaluate the performance of our system, relying on specific success metrics that align with individual clients. These metrics help us determine the reliability and efficacy of our system. One commonly used metric is the "add to cart rate," which indicates whether users add recommended items to their basket. The choice of metrics employed depends on the specific user or context of usage. For instance, in the case of search recommendations, we might measure how far down the page users scroll to find the product they eventually add to their basket. In the case of substitutions, we assess the "filled rate," which compares the number of accepted substitutions to the number of rejected substitutions. Once we determine the relevant metric, we establish real-time monitoring to track its performance. This allows us to identify and address any issues that arise promptly. Given the dynamic nature of our application, where models are continuously updated and modified, real-time alerting becomes essential. As soon as an anomaly or problem occurs, we receive immediate notifications, enabling us to take swift action. The effectiveness of our real-time alerting ensures the robustness and functionality of our system.

Grocery Doppio: What are some unique insights about customer behavior you have learned in improving personalization for grocery shoppers.

Michaelson: Through our work, we have realized the tremendous significance of product reorders in the online grocery shopping experience, surpassing our initial expectations. Strong reorder models play a foundational role in empowering various use cases. To enhance our clients' online grocery shopping experience, we prioritize optimizing and developing smarter reorder solutions that cater to individual tastes and preferences. This entails ensuring that when users search for items like arugula or lettuce, those items are prominently displayed at the top of their search results. By implementing robust reorder models, we can prioritize the products which shoppers frequently reorder, providing them with a seamless and efficient shopping experience.

Grocery Doppio: Who is the target audience of shoppers you engage with? How do you ensure shopper privacy while optimizing recommendations.

Michaelson: Our typical engagement is with shoppers with larger basket sizes, allowing them to maximize the benefits of our recommendations. These individuals are the primary beneficiaries of our platform. However, as our systems continue to evolve and mature, we also witness increased engagement from a broader range of grocery shoppers. We value and appreciate the growing interest and participation from a wider audience, as it allows us to serve and cater to the needs of a diverse customer base.

Nipote: In addition to engaging with shoppers with larger basket sizes, our focus extends to individuals who regularly utilize online grocery services. Similar to platforms like Spotify, the more frequently users engage with our system, the more they benefit from high-quality recommendations.  

Regarding demographics, it's important to note that one of our core principles is our commitment to not using personally identifiable information (PII) to power our algorithms. We take pride in this approach, which ensures user privacy and addresses various concerns associated with using PII.

Our strong adherence to the absence of personally identifiable information (PII) is a fundamental business principle that ensures the future scalability and adoption of our solution by retailers for many years to come. Unlike traditional approaches that rely on segmentation and demographic data, such as age or location, typically employed by coupon providers, our focus is more akin to Spotify. We concentrate on understanding shoppers' behavior and their interaction with food products and shopping patterns.

The target users who derive the most value from our platform are those who actively and frequently engage with it, thereby providing rich data that enhances the algorithm's effectiveness and making feedback loops to improve the system.

Grocery Doppio: How do successfully recommend products without using demographic information?

Michaelson: We have discovered that our models perform exceptionally well by focusing on shopping behavior and leveraging information from rich product features to predict future actions. As a result, we have yet to find it necessary to incorporate demographic features into our algorithms. Furthermore, we have taken privacy concerns into account, particularly concerning retailers. By avoiding using PII, including demographic data, we alleviate potential privacy risks and simplify the innovation process. Therefore, considering the current business landscape, we have determined that we need more than incorporating demographic information.

In many machine learning systems, the typical approach involves creating anonymous user profiles for each individual based on their previous shopping history. By analyzing their order patterns, these systems attempt to predict the products that users will likely purchase.

Initially, the system may perform poorly in predicting user preferences, but as it undergoes millions of training iterations, its performance gradually improves. The system generates individual user profiles through data generation techniques, which aim to approximate their future buying behavior and identify the products they are most likely to purchase. The critical factor we've observed for generating accurate user profiles is obtaining clean and consistent data.

Training models become challenging when there is missing demographic data for some users. It is easier to train models using a consistent and complete dataset, as this provides a more reliable signal for extracting rich features. Handling missing data poses difficulties in training and may affect the accuracy and effectiveness of the models.

Nipote: In a TV channel or similar platform, there may be considerations for targeting specific viewer demographics, such as the predominant viewer base falling within a particular range of age, gender, or ethnicity. However, it's important to note that our focus is on leveraging deeper levels of data to gain valuable insights. 

We believe that relying on demographic information was a strategy employed in the past to address specific gaps and capture customer preferences using broad categorizations. As our understanding and capabilities evolve, we aim to move beyond crude demographic buckets and delve into more refined and nuanced data analysis to provide personalized recommendations and experiences.

For instance, let's consider three individuals: Henry, Spencer, and myself. Despite having similar backgrounds, being of the same age, and growing up in the same place, we have distinctly different food preferences. This demonstrates that understanding someone's food preferences is more important than relying solely on demographic information. In our case, even though Henry, Spencer, and I share similar demographic characteristics, our unique food preferences highlight the personal nature of food choices. Our experience has shown that gaining insights into individuals' specific food preferences outweighs the predictive power of traditional demographic-based approaches.

Price: There's an additional aspect specific to the grocery industry that further illuminates the uniqueness of food preferences. Food is a uniquely nuanced category; in this category, shoppers have distinct relationships, limitations, and restrictions with the products they purchase. Moreover, grocery shoppers typically buy for an entire household rather than just themselves. This means that understanding a shopper's purchase behavior and what they're not buying or trying can provide valuable insights into the composition and makeup of their household.

This particular factor adds to the significance of providing a personalized recommendation and relevance engine in the grocery e-commerce experience. It's a level of relevance that may apply less effectively in other contexts. While demographic data may be relevant for driving traffic, when it comes to the point of purchase, the focus shifts to the end user and their specific relationship with the assortment of products. This unique dynamic gives us an advantage as a specialized vertical solution provider in the grocery industry, similar to how Henry described Spotify's approach to music. If Spotify had to consider how people interacted with phone cases, chargers, movies, or TV shows, it would present an entirely different challenge altogether. 

Grocery Doppio: What are some of the challenges you faced while implementing Halla's taste Intelligence Platform with your initial clients? What were the reservations they had?

Michaelson: At a high level, startups in the software industry often achieve success through rapid iteration. They can iterate quickly with a small group of beta testers and a few thousand users to work out any issues and reach their desired goals. 

However, in the case of large enterprise retailers, this process becomes challenging. 

The journey from having a minimal user base on the platform to successfully deploying it with their first customer and encountering a sudden influx of users is an immense and arduous process. It requires tremendous effort and perseverance. However, retailers often have high expectations for immediate return on investment or immediate benefits during the implementation phase of the platform and achieving them takes a lot of work.

Building software typically involves multiple iterations and refining with beta testers before achieving a stable and reliable product. This stark contrast poses difficulties for retailers, as the transition is more like opening a dam for a dry waterbed, with expectations of perfect performance right away. Such a scenario creates challenges in attracting retailers and generating excitement, as the risk associated with going from zero to a large user base is significant. It is unproven territory during the initial implementation.

From a technical perspective, there are considerations regarding scalability and defining the right metrics to optimize for when entering uncharted territory. It often requires learning from failures and experiencing multiple attempts with different retailers before achieving successful lift-off.

Nipote: Another challenging perspective is "prioritization." Selling our solution to these organizations is difficult, given the long list of priorities and the thin profit margins in the grocery industry. Operating a grocery business is an exceptionally demanding landscape, and introducing e-commerce adds an extra layer of complexity. It becomes even more challenging for retailers to invest significant time and resources in an implementation that may yield little benefits.

We have learned valuable lessons in prioritization during this process, making our solutions easier for retailers to adopt and providing them with a way to take the first step without undergoing a lengthy and risky implementation process. We had to ensure compatibility with the specific needs of grocers, which required iteration and refinement in the initial stages of our engagements. By offering a solution that aligns with their requirements and allows them to build at their own pace, we have found a way to address these challenges.

Grocery Doppio: According to our research, online grocers report that profitability is their number one problem; how can your company help increase digital profitability?

Nipote: This question is an extremely important one. To answer it, let me provide some initial insights before passing it on to Henry. Our solution has proven to significantly increase the rate of products being added to the cart, often resulting in a remarkable 10x boost in add-to-cart rates through recommendations. This presents a substantial revenue opportunity for grocers if they can successfully convert those added products into purchases and improve profitability. 

Increasing profitability is a crucial aspect, and one way to achieve this is by addressing the challenges online grocers face. According to a study by McKinsey, online grocers often experience a loss of $13 for every $100 order. Our solutions are specifically designed to help mitigate these losses, and our unique proposition is our focus on being grocery-specific, evident not only in our algorithms but also in the overall product design. 

One valuable lesson we've learned is the importance of collaboration between merchandising and e-commerce teams in creating an effective website. However, these teams often face misalignment in their goals and lack the necessary tools to unify their efforts. To address this, we have developed a platform tailored to address the detailed grocery requirements that enable seamless collaboration between e-commerce and merchandising, allowing them to work together and optimize operations and recommendations while achieving their respective goals.

Our platform includes tools and features specifically designed to support merchandising objectives, including profitability. Now, I'll hand it over to Henry to delve into further details regarding these features.

Michaelson: Recommendations play a significant role in driving profitability, as mentioned by Gabriel. Each additional product added to the cart increases the potential margin, making it more profitable to encourage customers to add more items. This is where Halla can assist by enhancing the add-to-cart rates and ultimately boosting revenue.

Another common aspect is for e-commerce and merchandising teams to have different priorities and sometimes conflicting perspectives. E-commerce teams focus on leveraging advanced technology, while merchandising teams aim to place specific products in the cart strategically.

We aim to integrate advanced technology and merchandising strategies seamlessly. We provide a real-time platform where merchandisers can utilize the advanced technology to promote specific objectives, such as increasing the visibility of private label brands by 15%. This integration ensures that recommendations align with merchandisers' goals and allows for targeted placements of products based on relevance and frequency, enabling the promotion of high-margin or high-priced items, boosting certain brands, and adjusting category priorities.

Halla aims to reconcile these differences by providing a platform with powerful personalization recommendation engines that allows both teams to collaborate effectively.

In addition to optimizing recommendations, we address the challenges of search functionality and substitutions in the online grocery experience. A well-functioning search feature is crucial for a positive shopping experience and the profitability of a grocery business. Practical search functionality is essential as it enables customers to find the products they need and build their baskets quickly. Studies have emphasized the significance of a good search experience, as it directly affects customer satisfaction and their willingness to spend money on the website. If the search feature is poorly designed or inefficient, shoppers may become discouraged and less likely to use the website, resulting in negative impacts on profitability and overall business success.

 Furthermore, substitutions pose another challenge for grocers. It has been observed that over 20% of customer orders may not be fulfilled as initially intended. This loss of basket value is significant. While implementing micro-fulfillment strategies can help mitigate this issue, it involves a complete overhaul of the existing business operations, which can be costly.

By addressing search and substitution challenges, we aim to empower grocers to build profitable businesses while delivering an exceptional online grocery experience to their customers.

The growth of online grocery needs to be explosive enough to warrant retailers scrapping their physical stores and transitioning completely to online operations. Given capital constraints and the ongoing evolution of online grocery, retailers will likely continue to fulfill orders through their stores. Consequently, out-of-stock situations are likely to persist. Our substitution engine, which leverages our recommendation and search engines, plays a crucial role in addressing this issue. It intelligently suggests suitable substitutions based on a shopper's cart, preferences, and previous shopping history.

This directly addresses profitability concerns by fulfilling a more significant portion of each order, leading to increased average order size. Moreover, minimizing the number of unfulfilled items is crucial, as failing to meet a shopper's expectations can result in losing that customer to a competing grocer. For legacy retailers, this poses a significant risk, potentially rendering their online grocery operations unprofitable.

Grocery Doppio: What are some of the grocer pain points you address in regard online profitability?

Nipote: Based on our experience with online grocers and consumer feedback, we have found that merchandising plays a crucial role in optimizing the online store experience, particularly regarding category pages. Throughout the entire shopping journey, merchandising significantly impacts the overall experience. 

One important aspect is providing relevant reorder recommendations on the homepage. However, we have encountered cases where category pages could be more organized, with products displayed in a confusing and unoptimized manner.

A common issue we have observed in online grocery stores is the suboptimal organization of category pages. For example, when clicking on a category page like "dairy and eggs," the products within that category may be displayed in disorganized fashion, with an overemphasis on certain items such as butter. It is crucial to address this issue as it impacts the entire website or online store. It's essential to recognize that traditional brick-and-mortar stores have spent years optimizing their layouts, strategically placing items like milk at the back of the store to encourage customers to explore other sections like bakery and produce.

 Another significant challenge in transitioning from traditional grocery stores to online platforms is replicating the strategies that drive impulse purchases in physical stores. Established practices to promote impulse purchases in conventional stores, such as placing high-margin items at eye level and utilizing checkout aisles for candy, magazines, and gum, contribute to approximately 20% of grocery sales. Adapting these tactics to the online environment is complex and poses a significant hurdle. The inability to capture this potential revenue can significantly impact profitability.

Another pain point lies in adopting personalization technology, as many businesses encounter challenges in integrating sophisticated systems that can effectively utilize customer data, conduct testing, and receive support for continuous improvement. Moreover, it requires close collaboration between merchandising and e-commerce teams to address operational integration.

Similarly, online grocers should prioritize effective merchandising techniques and personalization recommendation technology to create a shopping atmosphere that appeals to customers and guides them through the website seamlessly.

Ultimately, we focus on enhancing the customer experience and increasing the add-to-cart by addressing these pain points in the online grocery industry.

Grocery Doppio: Can you discuss any success stories where your personalization technology has led to increased sales or customer satisfaction for a grocer?

Nipote: We have achieved significant success with top-tier retailers, including those with annual revenues of $70 billion or more. We have also been deploying our solution with a range of grocers, including high-end, mid-tier, and lower-end top-tier grocers. Across multiple deployments, we consistently observe a remarkable 10x increase in the "Add to Cart" rate from recommendations, which is an exciting outcome. By suggesting more products, we contribute to the grocer's revenue potential by increasing the number of items in customers' baskets. Although our focus is not solely on conversion rates, the opportunity to boost revenue through higher cart values is valuable.

Grocery Doppio: Considering the fierce competition in the industry, what is your company's competitive advantage? What makes it stand out from others?

Michaelson: One aspect that sets Halla apart from others in the market is our strong focus on content-level understanding. We have developed a system that utilizes a homegrown ontology to organize the rich array of features associated with every product in a grocery store. This enables us to extract valuable information from the descriptive text of products.

Drawing from our initial experience building recommendation engines for restaurant dishes, we have acquired valuable knowledge that guides us in tackling the challenge of linking vast amounts of data. To address this challenge, we recognized that there is a limited number of commonly made dish types, such as burritos or sushi rolls, including variations like the California roll with different attributes. To process and analyze the descriptive text of products, we developed advanced technology capable of parsing this information.

For instance, if a customer is exploring a California roll, our system can identify similar California rolls across different restaurants, including variations like spicy California rolls. This required the development of a robust natural language processing (NLP) system that can extract relevant features from the product descriptions. These features encompass the product type (e.g., California roll), flavor profiles (e.g., spicy), and dietary considerations (e.g., vegetarian, chicken, or pork options).

In the context of grocery stores, the inventory can comprise approximately 100,000 products, generating tens of millions of associated events. This wealth of data provides valuable insights into product correlations tailored to specific user preferences. On the other hand, recommending restaurant dishes poses a different scenario, as the available data is typically limited. Each restaurant offers a unique menu with potentially millions of different products, making it challenging to gather important interaction data.

For example, if a customer searches for a "sesame bagel," our system can recognize it as a type of veggie top savory bagel, allowing us to provide relevant search results, substitutions, or complementary product recommendations. By leveraging NLP, we can link together similar products, enabling us to provide accurate and relevant recommendations for grocery items based on customer preferences and dietary requirements.

Many of my friends are huge fans of Spotify's Discover Weekly feature, which is truly remarkable. It serves as a primary means for discovering new music, and its recommendation engine operates similarly to ours. Spotify takes a proprietary and highly specialized approach to understand music by analyzing factors like the song's tempo and emotional tone, and they have developed specific algorithms to detect these features. They then combine this music analysis with real-time user behavior to generate a personalized set of recommendations.

Grocery Doppio: Please provide insight into any upcoming products or services your company plans to release.

Price: While we have yet to launch it publicly, I would like to provide a glimpse into an exciting opportunity. Our engine has the potential to respond to consumers' choices at the point of purchase, aiming not only to make their baskets bigger but better. This opportunity extends beyond driving incremental revenue. It has the potential to significantly impact the quality of life for consumers, particularly in North America, by helping them navigate toward a desired direction.

We recognize that individuals have unique lifestyle goals, health and wellness aspirations, nutrition objectives, limitations, and restrictions. By understanding these factors, we can guide and nudge their selections in a direction that aligns with their chronic conditions, diseases, health goals, or ambitions.

 We have made remarkable progress in developing this capability, which revolves around health and nutritional elements in the digital grocery experience. While we have yet to launch it, we are excited to power this initiative for our retail partners.

Michaelson: We recognize the critical health challenges faced by individuals in the United States and the impact their lifestyle and diet have on overall well-being. This presents a significant emerging market where we aim to make a positive impact by addressing people's dietary needs and improving their health. In doing so, we also provide an opportunity for grocers to increase their incremental revenues by catering to the growing demand for healthier options in the online e-commerce market in personalized nutrition.

Our primary goal is to assist individuals in developing personalized and manageable nutrition plans tailored to their specific requirements. Like skilled nutritionists, we understand the importance of listening to individuals, their preferences, unique tastes, dislikes, and dietary needs. While it's common knowledge that consuming vegetables, grains, and other nutritious foods is beneficial, we aim to go beyond generic recommendations by customizing our approach to each shopper.

At Halla, we believe that true transformation in the industry comes from customizing the grocery shopping experience for each individual. We aim to introduce unique product combinations and alternatives that cater to their tastes and align with their health goals. Unlike a mere filtering process, our approach is rooted in a deep understanding of an individual’s needs, desires, and circumstances. By offering personalized recommendations, we go beyond adding health features to create a grocery shopping experience that resonates with each shopper.

Grocery Doppio: How do you see the market for AI/ML-driven personalization in the grocery industry in the next 5-10 years, and how does the company's vision fit into this?

Price: We aspire to be at the forefront, like the sun, within the digital food solar system. Our objective extends beyond providing grocers with data and tools; we aim to illuminate their path and create exceptional real-time experiences. Our primary focus is enhancing the e-commerce landscape through innovative solutions such as recommendations, searches, and substitutions. As we look ahead to the next five to 10 years, we anticipate the emergence of various disruptive technologies in retail innovation, including digital signage, smart carts, scan and go, and frictionless checkout systems. By integrating these advancements with personalized nutrition-oriented assistance tailored to each individual, we foresee the potential to revolutionize people's perception of grocery shopping. Our ultimate goal is to shift the shopping experience from an unguided and monotonous routine to a guided, trusted, and exploratory journey. Through our commitment and contribution, we aim to play a significant role in shaping and facilitating this remarkable transition, particularly within the brick-and-mortar grocery industry, where our impact can be truly profound.

 Quotes that inspire

Grocery Doppio: Can you tell us an impactful piece of advice you followed during the challenging days that helped you continue your journey?

Nipote: "If you're going through hell, keep going." - Winston Churchill

 "If you don't know where you are going, you'll end up someplace else."

- Yogi Berra

‍

 "Your brand is what other people say about you when you're not in the room." 

- Jeff Bezos 

Spencer: "Diligence is the mother of Good Luck." - Benjamin Franklin