PKU families count the amino acid phenylalanine to the milligram, for life, and AI apps are starting to carry that work. This month, developers across the Claude hackathon are testing those tools with us. The first finding is clear: the AI can do the math, and the nutrition label is what holds accuracy back.
Developers gave us strong formulas. The reliable move is to have the AI sort each food into the right protein type first. Sort it right and the phenylalanine number follows by arithmetic, so the same food returns the same answer every time.
Working out how much of each ingredient sits in a bowl, from a short ingredient list, has no settled method. This is the problem we are handing to every developer in the hackathon for the rest of the month.
Families already bridge the gap between the label and the real phenylalanine by hand, with a gram scale and mental math. We are building a way a computer can repeat, so those logs count as clinical research evidence.
Open community testing runs all month, starting at the NPKUA meeting this week.
Open the repo, tell your agent to run
AGENTS.md, and walk away. It sorts 719 foods into protein types three times,
scores itself against USDA truth, and opens a pull request with the numbers. That is one
more data point on which models estimate phenylalanine reliably.
The reliability work above is the science spine. This is the other half: PKU as a market. Coverage law state by state, every medical food and formula and drug, the companies, apps, medical devices, clinics, dietitians, dental, mental health, restaurants, schools, trials, advocacy, and litigation. 433 cited records so investors and innovators can see where the opportunities are.
Most PKU apps read phenylalanine from the same place, and that place runs out of answers exactly where families need them.
PKU apps generally read phenylalanine from USDA FoodData Central. It has measured phenylalanine for single ingredients like an egg or an apple. It has almost no phenylalanine for packaged, composed foods like a candy bar or a frozen meal.
For those foods, the record is just the Nutrition Facts label: protein rounded to the whole gram, and no amino acids at all. So the app guesses, and different apps guess differently. The result: families, clinicians and researchers do not know if they can rely on app data.
USDA cannot fix this by measuring 350,000 packaged products. The change that scales is to calculate phenylalanine from the recipe, using measured ingredient values, and to say so on every number.
The same infrastructure serves three communities. Pick the door that fits you.
Contribute to the Claude phe-estimator Skill, the food list, and the scale integration. Get scored on the public benchmark. Ship apps that stay live.
See exactly how phe estimates are computed and validated, test Phebe on real labels, and tell us what a clinically-useful tool must get right.
Understand what these tools do, how we check they're reliable, and how your feedback shapes the diet-for-life tools your family relies on.
PKU (phenylketonuria) is an inborn error of metabolism: the body can't break down phenylalanine (phe), an amino acid in all protein. Left unmanaged, phe builds up and is neurotoxic. The treatment is a "diet for life," a very-low-protein diet that demands daily, gram-level precision.
Families can't use ordinary nutrition labels: protein grams are too coarse, and different protein sources carry different amounts of phe. So parents run an intuitive algorithm, deterministic math plus recipe-science judgment, to turn a label into a phe estimate, then weigh and log every bite.
A new wave of AI phe-estimator apps is taking the labor out of that grind. That's genuinely good news. But there is no standard by which a family can tell if an app is accurate, or by which a clinician can trust its output, and prototype results vary widely.
PKU apps also tend to go unmaintained. When a solo developer stops, the app stalls and families lose a tool their routine depended on.
Each is independently reviewable and independently reusable.
Encodes the "think like a PKU parent" method: is this food countable (<2 g protein/serving)? Which ingredients carry phe? What's the recipe factor? Multiply by portion. Log it.
A maintained low-protein foods reference, forked in part from Open Food Facts, that acts as a knowledge graph to raise reliability in the very-low-protein staple range.
Reverse-engineered gram-scale support so weighing flows straight into logging. Goal: connect the scale you already own instead of buying a new one.
The heart of PKU Commons is a public, reproducible benchmark. Every test case pairs a food label with a phe value computed from documented sources (USDA FoodData Central / Open Food Facts). Anyone can re-derive the "true" answer.
Any estimator, our Skill or any app implementing the interface, is scored the same way: error per case, mean absolute error, and percentage of estimates within a clinically-relevant tolerance band.
One benchmark, three jobs:
PKU Commons answers to two traditions at once. Clinicians and researchers ask "is this defensible?"; engineers and families ask "does it measurably work?" Those map onto two review methods. We use both, on different parts of the system.
Every estimation rule and every food-list row carries a citation to authority (a USDA FoodData Central id, an Open Food Facts code, or a named clinician / dietitian sign-off), a version, and a reviewer of record. Challenges are filed as issues; resolutions become part of a permanent audit trail. The record shows what is in the knowledge base, why, on whose authority, and who can be asked.
Accuracy claims are settled by measurement against the public benchmark, independent of argument. Ground truth is reproducible from documented sources. Every estimator is scored the same way. Contributions merge only when they improve the score without regressing, enforced in CI. The benchmark above is this layer in action.
NPKUA's TruPKU / Voice of the Patient report catalogs the friction of living with PKU beyond the clinic and the lab. Each pain point below is a suggested hack for the PKU Commons community, a concrete place to point AI and devops at the real burden of living with PKU. Themes summarized from the public report; see pain-points.mdfor the source of record.
Every food is weighed, calculated, and logged by hand. It becomes a tedious, never-ending accounting chore that families eventually abandon, and the clinic loses the data.
Hack: Cut estimation-to-log to one tap: label parse, scale weight, and auto-log, verified against the benchmark so families can trust it.
Blood phe results come back 5 to 14 days later by mail. There is no timely link between what someone ate, how they feel, and their measured level.
Hack: Build the food, symptom, and level data link: structured logs ready to correlate as soon as faster home testing arrives.
USDA protein grams are too coarse. Phe per gram varies by protein source. Families reverse-engineer every label by hand.
Hack: Improve the ingredient-to-phe classifier and grow the living food list so the very-low-protein staples are covered and cited.
Apps built by one parent or one small shop go unmaintained and disappear, leaving the families who relied on them without a tool.
Hack: Ship on shared, benchmark-gated infrastructure so an app keeps working and stays trustworthy after its author steps away.
Estimation happens in a grocery aisle on weak store Wi-Fi, one-handed, under time pressure, far from a lab.
Hack: Design for the aisle: offline-tolerant, low-bandwidth, and deterministic behavior, with graceful degradation when the network is poor.
Powerful new PKU therapies can over-correct into hypophenylalaninemia, which needs a fast dietary response, but the dietary data loop is slow.
Hack: Make dietary phe data fast and shareable enough to support quick, safe adjustments alongside modern treatment.