Course:FNH200/Projects/2025/Interview with Honeycomb AI CTO

1. Inspiration Behind the Creation: Motivated by personal experiences, his co-founder has dietary restrictions, and he himself has a brother with a severe fish allergy. These challenges sparked the idea of creating a solution for people with similar needs. As he explored the idea further, he became fascinated by the complexity of food data, which deepened his interest in pursuing the project.
2. How nutritional accuracy in models ensured: The interviewee explains that the model primarily uses text-based data rather than images. While images of menus can be provided, the text is what’s actually analyzed. They have a large pre-trained dataset of labeled restaurant items that helps in making predictions. The model’s predictions are always verified by the restaurant, ensuring that there’s a human-in-the-loop to catch any inaccuracies. The accuracy is around 90%, but the final check by the restaurant ensures that any errors are corrected, making the process efficient and reliable. The goal is to act as a co-pilot for the restaurants rather than fully automating the process without oversight.
3. Interesting facts or trends discovered in the food industry: One surprising fact is that many restaurants often underreport their calorie counts. They rarely conduct lab tests for each dish, which leads to discrepancies between reported and actual calorie content. Another interesting trend is that while restaurants try to be transparent, sometimes even their own staff might not know all the ingredients in a dish, especially in more complex or specialty items. These discrepancies are more common in mid-sized restaurant chains that have multiple locations but don’t have the resources or legal requirements to conduct rigorous testing, unlike large fast-food chains that are more strictly regulated. Another fascinating insight was that Honeycomb can often predict the "secret" ingredients in dishes — a capability that challenges the idea of truly unique recipes, as it suggests AI can demystify even closely guarded culinary secrets.
4. Handling inaccuracies in calorie predictions, especially with images: The interviewee explains that it’s challenging to predict calories accurately just from images because portion sizes can vary. The models might be good at identifying what type of food is in the image, but not how much of it there is. The interviewee mentions that more advanced technology, like LiDAR sensors, could potentially improve accuracy by measuring portion sizes more precisely using depth as the metric, but it’s still a developing area. Overall, the current models rely on a combination of text and image data, but they acknowledge that predicting calories from images alone isn’t fully reliable yet.
5. Rules or laws related to food in honeycomb app that need to be followed: They do have to make sure they’re legally covered. They include disclaimers to let users know that the nutritional information is a recommendation, not a lab-verified fact. There’s an emphasis on liability protection, and the app’s interface also indicates to users that some variation or error can occur. Ultimately, the restaurant verifies the information, and that adds an extra layer of trust and accuracy.
6. Retraining models and evaluating user feedback in order to improve them: They do incorporate user feedback into retraining the models, especially when users point out discrepancies in calorie counts or nutritional details. Some users are very detail-oriented and help improve the data, while others are less precise. They have to balance between these different types of feedback and the overall data quality to account for personal bias. Overall, user input is a valuable part of refining the model and making it more accurate over time.
7. Consulting with professional food scientists or consultants to improve model accuracy: Yes, they do consult with food scientists and nutritionists to guide them in improving and retraining their models. They also have a doctor who specializes in food and health on their advisory team, which helps add credibility. While they rely heavily on their own data sources, consulting these professionals definitely increases the credibility and reliability of the model.
8. Trends related to ingredient transparency and consumer trust: While they don’t have specific data directly linking ingredient lists to sales, there is a general trend that shows that greater transparency can build more trust with consumers. The interviewee mentions that platforms like Yuka or the Daylight app show that when brands are more transparent, it often leads to a stronger relationship with users. Beyond that general observation, they haven’t deeply analyzed the correlation, but transparency does seem to have a positive impact
9. Feature in development that addresses cultural dietary needs like halal or kosher: Yes, they do currently handle this at the restaurant level. If a restaurant is certified halal or kosher, that information is included so users can see it. At the ingredient level, they don’t have that feature yet, but it’s definitely something they’re considering adding in the future based on user interest.
10. Working towards global outreach: Model is best used for English speaking languages but is built for global use. They are currently working on expanding into new regions like the Middle East and India, collaborating with companies in those areas to gather and incorporate local data. They’ve also had interest from Australia, noting that while there are similarities with North American cuisine, there are unique ingredients and culinary traditions in each region that they need to consider. The interviewee also mentioned that global outreach is an ambitious feat to achieve since countries outside of North America use unique spices and ingredients and accounting for those in Honeycomb requires rigorous research and involvement with the community.
11. Other challenges besides getting accurate calorie information: Another major challenge is data acquisition. Gathering comprehensive and reliable data from different regions and cuisines can be difficult. The data varies widely, and ensuring consistency and accuracy across different datasets is essential but complex.
12. Biggest advantage of Honeycomb that gives it a competitive edge: One of the main advantages is that Honeycomb is very user-focused. They strive to understand the problem from the user’s standpoint and tailor their solution closely to what users actually need. Another big advantage is their AI capability. The interviewee notes that food tech companies are often a bit slower to adopt advanced AI, and Honeycomb’s models are ahead in that space, achieving higher accuracy where others might not.
13. Data collection strategy for the model: At a high level, they start with publicly available data, such as restaurant menus. They also have partnerships that allow them to buy or access more specialized data. For instance, they partner with other businesses that provide detailed data on a large number of restaurants.In addition to external data, they have their own internal labeling and verification processes, using AI models to help refine and ensure the accuracy of the data.
14. Looking into solutions for ingredient storage and offering different storage options based on ingredient types: While ingredient storage solutions are interesting and definitely a recognized problem, it’s not currently a focus for Honeycomb. They’re aware of the challenge, but it’s not something they’re actively working on at the moment.
15. CTO of a food tech company's main responsibilities: The CTO’s main responsibilities include product and model development, focusing on the nutritional models and the technology behind them. They also play a key role in enterprise sales and partnerships, ensuring that the product can integrate well with other systems. Additionally, data management and accuracy are key parts of the role.
16. Company size: The company is still relatively small, with around 20 team members, but it’s growing steadily. They have formed partnerships with well-known restaurants and food brands, and there’s a lot of room for future growth.

Connections drawn to the course:

Connection 1: Food Science Category Focus - Food Analysis

Key questions asked: What does Honeycomb.ai do? In terms of food analysis, how does the company go about collecting data? (2:09 - 3:40)

In FNH 200 lesson 1: Food Science and the Canadian Food System, we learned about the intra-disciplinary diversity of food science and its major sub-fields. As pertains Honeycomb.ai and its operational focus on obtaining, analysing, and disseminating data on key allergens, ingredients, and dietary factors within dishes, the company’s innovative concept and approach discernibly place them within the food science category of “Food Analysis,” which, as described in the lesson, concerns “principles and methods for quantitative physical and chemical analyses of food products and ingredients.” As the CTO explained, the model through which Honeycomb.ai acquires its data is predominantly based on textual information provided by restaurant menus, which is then verified by the restaurant in question in order to mitigate and address potential inaccuracies. Furthermore, Honeycomb.ai regularly consults with food scientists and nutritionists to assist in the betterment and retraining of their models, in addition to having a food and health professional on their team, both of which greatly adds to the company’s credibility by demonstrating their commitment to processual rigour and due-diligence.  Of import to note was the CTO’s explanation of the vast potential presented by LiDAR sensors to capture data through a depth-based imaging methodology which may increase accuracy through measuring portion sizes more precisely. Ultimately, Honeycomb.ai’s pioneering approach illustrates the burgeoning importance of AI for the future of food science.

Connection 2: AI Integration (3:40-4:28)

Key questions asked: How is Honeycomb.ai showing the effect of AI in the food industry? What is the model’s success rate and how can reliability be ensured?

IFT’s Top Ten Food Trends for 2025 predicted a rapid expansion of AI integration in the food industry—a trend clearly reflected in Honeycomb’s recent success. As online food ordering continues to grow, it's only natural that individuals with dietary restrictions or fitness goals want tools that cater to their needs. Yet, limited customization options have often stood in the way. Honeycomb addresses this gap by using AI to provide tailored recommendations based on allergens and calorie tracking, boasting an impressive 90% accuracy rate. This alignment between Honeycomb’s innovation and the broader AI trend in food is unmistakable.

Connection 3: Food related hazards/Accommodating to dietary restrictions (14:36-15:38)

In FNH 200, Lesson 12 introduced us to the concept of food-related hazards and their impact on human health. Honeycomb contributes meaningfully to this space by identifying key allergens in food products, enhancing the safety of online food ordering. Where users once had to guess or rely on vague descriptions and reviews, they can now access real-time, data-driven insights verified by restaurants themselves. This reduces the risk of adverse reactions and makes online food choices safer and more informed.

Connection 4: Standardization and Food legislation (10:06-11:24)

In FNH 200, Lesson 4: Food standards, Regulations, and Guides, we were introduced to and familiarised with the concept of food regulation and standardisation (within the context of Canada), in addition to the primary government agencies concerned with regulating food quality and safety. Upon discussing the legal parameters to which Honeycomb.ai is subjected to through its operations, we learned that liability regarding their dissemination of nutritional content is mediated through their full transparency with users that such information is a “recommendation, not a lab-verified fact,” in addition to maintaining an interface with users which discloses the potential occurrence of errors or variation in data to the latter. The legal aspect is ultimately the responsibility of the restaurant/distributor, whose verification of Honeycomb.ai’s data provides an added level of both trust and accuracy. Honeycomb.ai’s recognition of the necessity to adhere to the law regarding information distribution concerning nutritional information reflects the importance of Canada’s emphasis on food protection laws, which is enshrined through the Food and Drugs Act, a piece of legislation comprehensively delineating the parameters around labelling and transparency of ingredient contents to which distributors much adhere to.

Potential Exam Question

Question: Given a prepared dish in a restaurant, what is the best way to predict the number of calories?

A. Upload the image to a LLM such as GPT

B. Use a custom pre-trained neural network model for prediction

C. Use LiDAR sensor to leverage depth and machine learning for prediction

D. Make an educated guess based on previous knowledge obtained by eating similar meals

Correct Answers: C

Explanation: Many current calorie-prediction projects suffer from high error rates because standard images lack depth information, making it difficult to estimate portion sizes accurately. LiDAR sensors, which measure depth using laser beams, can precisely determine the volume of food on a plate. When combined with machine learning models—which are already effective at identifying the types of ingredients—this additional depth data allows for much more accurate calorie estimation. While machine learning alone can classify foods and its ingredients, it struggles with quantifying them, a key component to calculating calories; LiDAR solves this limitation, enabling significantly improved predictions.

Why this question should be on the final exam: This question addresses a real-world challenge faced by the CTO of Honeycomb AI and other food technology companies. With AI applications expanding rapidly in the food industry, students of FNH should understand both the limitations of current AI methods and how integrating complementary technologies like LiDAR can overcome them.