From: aidotengineer
Alma believes that eating well should be simple and good health begins with personalized nutrition [00:10:10]. Traditional nutrition apps often require users to share significant information while providing little value in return [00:34:04]. Alma’s vision is to leverage AI to overcome these challenges, creating a “true nutrition companion” that helps users achieve their best selves [00:44:03].
Core Pillars of Alma’s AI Nutrition Companion
Alma’s approach to personalized nutrition through AI is built on three main pillars:
1. Simple Nutrition Tracking
The goal is to make nutrition tracking easy, natural, and intuitive, similar to texting a friend [01:04:47]. This avoids the laborious process of sifting through extensive product lists or relying on inaccurate photo recognition [01:13:00].
2. Contextual Understanding
Once tracking is easy, the information gathered is used to build a robust context around the user [01:23:14]. This includes understanding their flavor profile, interests, habits, and hobbies to guide them towards their stated goals [01:29:08].
3. Personalized Recommendations
The ultimate vision, set to launch in the second half of the year, involves using collected data to connect users with specific products, restaurants, and meals that align with their health objectives [01:36:38].
The Alma Score
Through user conversations, Alma discovered that users desire a holistic understanding of their food quality beyond just calories and macros [02:14:04]. Recognizing that not all calories are equal in quality [02:24:26], Alma developed the “Alma Score” concept in collaboration with academic adviser Dr. Eric Prim from Harvard University [02:30:57].
The Alma Score is a simple rating out of 100, designed to guide users towards foods scientifically proven to be beneficial for health and away from less healthy options [02:37:37].
Lessons Learned in Building an AI Nutrition Product
Alma has gained valuable insights from building its AI platform:
User Feedback is Paramount
Relying on user feedback is more important than evaluation metrics, especially in the early stages of AI product development [03:01:21]. Alma implemented a “How did Alma do?” prompt after every interaction to gather real-time feedback, which proved invaluable for guiding development and measuring accuracy [03:17:42].
Constraining LLMs for Accuracy
Large Language Models (LLMs) tend to have high error rates when given broad, open-ended tasks [03:39:09]. It’s crucial to constrain the LLM to very specific tasks [03:48:00]. Alma achieved this by breaking down the information relay process into steps. For example, when a user says “I had a banana,” the AI first extracts “banana” as a food item, sends it to the client, and then processes it through a matching system with the USDA database to find caloric content [04:06:09]. This step-by-step approach makes the experience faster and more interactive for users [04:46:08].
Gamification (Streaks) Matter
Features like streaks, initially considered minor, proved to be highly valued by users [05:03:07]. An accidental break in streaks caused significant user uproar, highlighting the importance of such engagement features [05:06:02].
Continuous Context Building
When building AI agents, it’s vital to continually build context about the user to prevent frustration from repeated information [05:39:00]. Alma maintains a “knowledge data set” about each user, automatically adding novel information from interactions, making the AI smarter over time [05:58:19]. Users can view, add, or delete this information [06:06:03].
Proactive Engagement and Insights
Users won’t always proactively open the app to ask questions [06:16:04]. Alma addresses this by proactively detecting insights from user eating habits and surfacing new information [06:30:04]. Examples include nutritional facts, such as how pairing blueberries with dark chocolate increases vitamin C absorption [06:36:39]. This creates a continuous, bite-sized learning process for users [06:54:19].
Embrace Multimodality
Initially, there was a focus on identifying a “winning” modality (e.g., voice) [07:01:05]. However, user engagement showed a preference for multimodality – the ability to interact via voice, photo, or text depending on the context [07:27:07]. Providing various convenient interaction methods is key [07:44:03].
Future Focus for Alma
Alma’s forward-looking strategy for AI in personalized nutrition is centered on three key learnings:
- Brand and Design: Code is becoming commoditized, so user-centric product design and appealing visuals are critical for standing out [07:59:03].
- Trust and Partnerships: Earning user trust, especially as a new entity, involves partnering with trusted sources who have an aligned mission [08:26:01]. Alma actively asks users about their trusted information sources to explore partnerships [08:36:06].
- Community and Net New Data: As large LLM companies aggregate vast amounts of online data [09:19:07], it’s essential to leverage users and community to generate valuable, net-new data [09:30:03]. Users are curious about others’ eating habits and potential learning opportunities [09:08:58]. This focus on community-driven data is a key area for Alma [09:16:08].