AI Engineer World's Fair 2025
My AI Thinks I'm Eating My Feelings (and Other Nutritional Insights)
Loading slides…
Read the talk
My AI Thinks I'm Eating My Feelings (and Other Nutritional Insights)
Alma’s nutrition companion connects simple meal logging with progressive results, user-controlled memory, and timely guidance. Its development shows how AI reliability and product design depend on each other.
From a talk by Rami Alhamad
Making nutrition tracking worth the effort
How much information should you have to give a nutrition app before it tells you something useful about your eating? Searching for foods, identifying products, and maintaining a detailed log can demand more effort than the resulting guidance justifies. Drawing on his experience in wearables, health, and wellness, Rami Alhamad introduces Alma as an attempt to improve that exchange: simple, personalized nutrition that makes eating well easier. The lessons come from roughly eight months of building the product.
The product vision proceeds through three connected capabilities:
- Make tracking natural. Logging a meal should feel like texting a friend, without searching endless product lists or relying on a photo and hoping the interpretation is close enough.
- Build useful personal context. Combine the log with a person’s flavor preferences, interests, habits, and hobbies to guide them toward goals they choose.
- Connect guidance to food choices. Recommend products, restaurants, and meals that help the person act on those goals. In the talk, this discovery work is planned for the second half of the year.
Easy input supplies the information that personalization needs; food discovery gives that personalization somewhere practical to lead.
Alhamad and co-founder Adam favor building in public and learning from users early. Alhamad reports that Alma shipped its beta less than two months after incorporation, spent roughly four months in closed beta, and launched publicly in February 2025. The company’s launch announcement dates that public release to February 5.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Learning what users need to know
Once people began using Alma, their questions extended beyond calories and macros. They wanted to know how well they were eating overall. A thousand calories can come from very different foods; the energy total alone does not describe dietary quality. Alhamad describes Alma Score as a score out of 100, developed with Harvard academic advisor Dr. Eric Rimm. Its intended function is to steer choices throughout the day, encouraging more foods associated with good health and fewer of those associated with poorer health. The talk describes that purpose, without providing a scoring formula or validation results.
The same need to understand users shaped the team’s approach to evaluating the AI. Alma had explored evals, but Alhamad found immediate user reactions especially useful during the beta. After every interaction, a drop-down toast asks, “How did Alma do?” That puts the feedback request next to the result the user has just experienced, helping the team identify problems and guide improvements as people track their meals. The team planned to reduce the prompt’s frequency after this early learning phase.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Show the banana before the calories arrive
Alhamad’s next lesson concerns task size. In Alma’s experience, giving an LLM a large, open-ended assignment created too many opportunities for error. The team responded by constraining models to specific tasks and designing the interface around that decomposition. A multistep pipeline can return useful partial results before the entire job finishes.
The example is a simple message: “I had a banana.” Alma processes it in stages:
- Recognize the food. A module identifies banana in the phrase and extracts it as a food item.
- Display the item. The extracted food goes to the iOS client immediately. The meal timeline can show Banana while other fields remain placeholders.
- Enrich the entry. Subsequent processing includes matching against a USDA database to find appropriate caloric content, which reaches the client later.
The important boundary is between knowing what food was mentioned and having its nutritional information ready. USDA matching is part of the later work; the talk does not identify the dataset, endpoint, or matching algorithm.
A TypeScript representation of this delivery pattern separates recognition from enrichment. The nutrition lookup stays behind a function boundary, while each completed stage produces an update for the client:
typescript
type Food = { id: string; name: string };
type Nutrition = { calories: number };
type MealUpdate =
| { stage: "recognized"; food: Food }
| { stage: "enriched"; food: Food; nutrition: Nutrition };
async function* trackMeal(
text: string,
extractFood: (text: string) => Promise<Food>,
matchNutrition: (food: Food) => Promise<Nutrition>,
): AsyncGenerator<MealUpdate> {
const food = await extractFood(text);
yield { stage: "recognized", food };
const nutrition = await matchNutrition(food);
yield { stage: "enriched", food, nutrition };
}
For the banana input, the first update lets the interface acknowledge the recognized item while calories are still pending. The second enriches the same food entry. Alhamad reports that this makes the interaction feel faster and easier than waiting for every processing step to finish before displaying anything; he provides no latency measurement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Preserving progress and personal context
A broken feature supplied another kind of feedback. The week before the talk, an engineer broke streaks, prompting a surge of complaints in Alma’s WhatsApp community and Intercom. Users wanted their streaks back. The team had treated streaks as a small, fun addition made toward the end of development; the reaction revealed how much users valued them. Alma consequently planned to invest more in the feature, keeping enjoyment part of the experience for both users and builders.
Continuity also matters inside the agent. Users become frustrated when they have to repeat information during meal tracking or when asking questions. Alma therefore extracts novel information from interactions and adds it to a user-specific knowledge dataset. A visible orb signals an update. Users can inspect the stored information, remove it, or add more themselves. Persistent context is useful when the person it describes can see and control it. Here, getting smarter through interaction means accumulating usable personal context; the mechanism described is a knowledge dataset.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Reach out, and let users choose how to respond
A companion cannot depend entirely on users opening the app to ask questions. Alma also detects possible insights in what someone eats and proactively surfaces them through pings. Alhamad’s example is an insight claiming that pairing blueberries with dark chocolate increases vitamin C absorption. That specific absorption claim is unsubstantiated here and should not be treated as established nutrition advice. The product mechanism is to turn food logs into timely opportunities to learn, extending the interaction beyond deliberate queries.
That bite-sized learning experience leads into another design question: which input modality should dominate? Alhamad initially favored voice. Alhamad reports that, when busy, he can log breakfast, lunch, and dinner together by voice at the end of the day in under ten seconds, compared with minutes previously. This is his personal logging experience, not a measured end-to-end processing benchmark.
Users changed his view of the question. They valued being able to talk, take a photo, or type depending on their circumstances. The lesson is to support the modalities that make sense for the user’s situation, rather than choosing one winner based on the founder’s preference. A natural interaction is partly a matter of having the right input method available at the moment someone wants to log food.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Design, trust, and data people create together
Looking ahead, Alma’s priorities begin with brand and design. Alhamad’s premise is that code is becoming a commodity, making a user-centered, visually appealing product increasingly important to standing out. Adam leads this design focus, which Alhamad describes as central to how the company intends to build over time.
The second priority is trust and partnerships. A new product needs to earn confidence, and Alma’s approach starts by asking users where they get information and whom they already trust. Those answers guide the search for partners with aligned missions and a mutually useful reason to collaborate. The partnership strategy follows users’ existing sources of confidence.
The third priority is community and new data. Food is social: people eat together, talk about meals, and become curious about how others eat. Alma members want to learn from one another’s choices and potentially adopt useful habits. That interest motivates planned community features and creates an opportunity to collect information that did not previously exist.
Alhamad connects this community opportunity to the economics of foundation models. His view is that large model companies probably already have much of the information a startup can scrape from the public web. Members creating valuable new information together offer a more distinctive source of data. For Alma, the strategic question becomes how to help people learn from one another’s eating habits while creating something useful that a public-web scrape would not contain.
The talk closes with an invitation to try Alma and join its WhatsApp community. Alhamad offers the audience the promo code ALMAFAM as part of that invitation; it is a historical offer from the recording.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Alma's iOS nutrition app supports meal logging through voice, photos, text, and barcodes, with personalized dietary guidance.
Further reading
The company's public launch announcement describes its initial positioning, input modalities, and scientific advisory relationship.
- USDA FoodData Central documentationDocumentation
A guide to USDA food-composition datasets, their sources, and the differences between analytical, survey, branded-label, and historical data.
- Vitamin C absorption and nutrition evidenceDocumentation
NIH's evidence summary explains vitamin C absorption, food sources, physiological roles, and interactions.
Updates since the talk
- Explore Food announcementArticle
Rami Alhamad introduces personalized restaurant recommendations, recipes, products, and community meals in Alma's Explore Food feature.
Full transcript
Loading transcript…