Alma makes food tracking feel like texting a friend by accepting voice, photos, and text instead of forcing users through long product lists.
2
The team reduces errors by splitting food tracking into small tasks, then streams results while later steps match foods against the USDA database.
3
User feedback has shaped Alma's product, from its per-interaction rating prompt and persistent user context to streaks and proactive nutrition insights.
Summary
Rami Alhamad describes Alma's attempt to build an AI nutrition companion that makes food tracking simple and turns meal data into personalized guidance. The product accepts natural-language text, voice, and photos, then processes a meal through separate steps rather than asking one large language model to handle everything at once. Alma also builds a user profile from novel information, gives users control over that stored context, and sends short food insights proactively. The team learned from its beta that users wanted a view of food quality alongside calories and macros, which led to the Alma Score, a score out of 100 developed with Harvard adviser Dr. Eric Prim. Alhamad is candid about product lessons: live user feedback was more useful than relying only on evaluations, streaks mattered more than expected, and users preferred multiple input methods. Looking ahead, Alma is focusing on design, trusted partnerships, community, and new data from its members.
Alma starts with low-friction tracking and personalized guidance
Alhamad describes three parts of Alma's nutrition companion. Tracking should feel natural, like texting a friend, instead of searching through product lists or trusting an imperfect photo result. The system then uses what a person records to build context around their flavor profile, interests, habits, and hobbies. That context should help guide the person toward goals they have chosen. Alma also plans to connect users with products, restaurants, and meals that fit those goals. The product vision therefore begins with easier input and extends into advice and recommendations.
The Alma Score adds food quality to calories and macros
After launching its beta, the team found that users wanted more than calorie and macronutrient totals. They also wanted a rough sense of how well they were eating overall. Alhamad points out that 1,000 calories can come from many different combinations of food, and those combinations do not have the same quality. Alma developed the Alma Score with its academic adviser at Harvard University, Dr. Eric Prim. The score is out of 100 and is intended to steer people toward foods supported by nutrition research while nudging them away from foods the team considers less healthy.
Real-time user feedback guides the product better than testing alone
Alma added a small drop-down prompt after every interaction that asks, "How did Alma do?" Alhamad says the team expects to reduce how often it appears, but found the feedback especially useful during the early product phase. It helped the team decide how to build Alma and measure changes in accuracy. He says the team has experimented with evaluations, yet direct feedback from users in the moment has been more useful for understanding how the system performs in practice.
Breaking food tracking into small model tasks makes the app faster
Alhamad says open-ended, very large tasks produce high error rates from language models. Alma works around that limit by dividing the response process into steps. In his banana example, one module recognizes the word in the user's message, extracts it as a food item, and sends that result to the iOS client. Other steps continue processing the item, including matching it against the USDA database to find appropriate calorie information. Sending partial results to the client makes the app feel faster and gives users something to interact with before the full process finishes.
Small product features can reveal what users value
An engineer accidentally broke Alma's streak feature, and the reaction from the WhatsApp community and Intercom was immediate. Users asked what had happened to their streaks and told the team not to change them. Alhamad had treated streaks as a small feature that could be added near the end, but the response showed that users cared about it. The team now plans to invest more in the feature and keep the product enjoyable for both the builders and the people using it.
Alma stores new information about a user in a visible area labeled "about you." When an interaction contains a novel detail, the system adds it to the user's knowledge set. People can inspect, delete, or remove those details. This lets Alma remember information across tracking and questions, so users do not have to repeat the same preferences or facts. Alhamad presents this continuing context as part of how the agent becomes more useful over repeated interactions.
Proactive food insights give users short lessons between questions
The team found that users would not open Alma often enough if the product only waited for questions. Alma can detect patterns in what someone eats and send information the person may not know. Alhamad gives the example of learning about blueberries and dark chocolate in relation to vitamin C absorption. These messages make food education a recurring, short interaction rather than a long research task. He describes the result as a continuing process of learning about food in small pieces.
Users want to switch between voice, photos, and text
Alhamad initially thought voice might be the best input method. He says he can turn it on at the end of the day and record breakfast, lunch, and dinner in under 10 seconds, whereas the old process took minutes. User behavior changed his view. People liked being able to talk, take a photo, or text depending on the situation. Alma therefore focuses on offering several modalities rather than choosing one input method for everyone.
"I've learned that you really want to provide your users with as many different modalities that make sense for them."07:25
Who should watch
You are building an AI product where users need to provide structured information through everyday conversation, photos, or voice.
Your model performs poorly on broad tasks and you need a practical way to split processing into smaller steps while returning partial results quickly.
You want product lessons about persistent user context, direct feedback prompts, feature surprises, and proactive engagement.
