Nutrition AI Agent · In public testing

Decode every meal.

Delemma is an AI agent for nutrition. It turns your food, biometrics and goals into the next decision — automatically. No labels to read. No spreadsheets to keep.

Apply for external trial → How it works

291k

Training images

647

Dish classes

Nutrients tracked

Goal modes

Origin

It started with a single number.

Delemma began as one person debugging his own metabolism. The principles that worked turned out to be the principles that work for everyone.

In 2023, mid-thesis, jet-lagged, and chronically underslept, I was diagnosed with a runaway A1C of 14%. The clinic's plan was lifelong insulin.

I went home and started treating my body like a system to debug. Every meal logged. Every biometric tracked. Every recommendation cross-checked against the literature.

Sixteen months later A1C was 5.5% — back inside the normal range. The thing that worked wasn't a diet. It was a feedback loop: data in, decision out, signal back.

"What works for one person who's paying attention works for anyone — if the loop runs fast enough."

Rocky Li · Founder, Delemma

A1C · 21-month trace

Built for everyone in motion

Three goals. One agent.

Pick the goal that fits today's life. Delemma rewires food, biometrics and recommendations around it — and adapts as the goal changes.

Maintain & optimize

Stable life, modern diet — but tired, foggy, sleeping poorly. Delemma maps 49 nutrients against your real intake and surfaces the silent gaps: vitamin D, magnesium, omega-3, B-complex.

EnergySleepMoodSkin

Cut without breaking

Calorie deficits collapse without protein, magnesium, B-complex. Delemma separates "calorie gap" from "nutrient gap" so you keep your metabolism — not just your scale number.

Protein floorMacro splitsBMR guard

Fast with precision

16:8, 18:6, OMAD — the upside lives in the eating window. Delemma calculates exactly the protein, potassium, magnesium and sodium your window needs, and pre-stages electrolyte protocols.

Window mathElectrolytesInsulin sens.

Delemma Flux · the agent layer

An AI agent fed by
your real-time data.

Flux is the conversational core of Delemma. Every reply pulls live signal — biometrics, intake history, energy balance, your goal — and answers in the moment, not in the abstract.

Live biometrics. Heart rate, sleep, blood glucose, SpO₂ — pulled from Apple Health, Apple Watch, CGM.

Today's plate. The full breakdown of what you ate, and the 49-nutrient bank against your goal.

Seven-day trend. The nutrients you keep underconsuming, and the meal patterns that keep producing them.

Energy ledger. BMR + NEAT + activity + thermic effect of food, reconciled against the day's intake.

The Stack

Owned, end to end.

Every layer of the agent is built in-house. The vision model, the nutrient ledger, the live data pipes, the nutrition-specific reasoning — none of them rented.

Vision · self-trained

A vertical food-recognition model trained from scratch — not a wrapper around GPT-4V. Marginal cost per inference rounds to zero, which is what makes a free tier sustainable at scale.

291K

Training Images

647

Dish Categories

Cuisines Covered

75.2%

Overall Test Pass Rate

Cuisine Cat. Samples Pass Rate

Chinese232~151K

100%

Japanese115~14.4K

100%

French26~7.0K

87.5%

Italian26~7.8K

75.0%

Mexican18~6.5K

75.0%

American74~62K

73.3%

Other Asia + EU + etc156~48.2K

Expanding

Total647~291K

75.2%

Ledger

Nutrients tracked · target vs intake

Domain reasoning · in training

A nutrition-specific model.

General medical models fixate on diagnosis and pharmacology. Nutrition is a different discipline: absorption sites, bioavailability, dose-response, hour-scale metabolic loops, food-supplement-drug interactions. We're training the model that should have existed all along.

Metabolic pathways DRIs / EFSA / WHO Multi-timescale reasoning Interaction graph

Live signal

Biometric pipeline

Continuous streams from the devices already on your wrist or arm — fused into a single live signal Delemma reasons against in real time.

Apple Health Apple Watch HealthKit CGM (Dexcom · Libre · Lingo) Smart BP monitor

Sustainable AI

85%+ less energy. Zero marginal cost.

We believe technological progress shouldn't cost the Earth. Delemma is engineered from the ground up as an eco-friendly AI agent. By optimizing our inference architecture, we cut energy consumption drastically compared to monolithic LLMs, allowing us to offer a sustainable free tier for everyone.

Small & Focused

Specialized SLMs

Replacing 175B+ parameter monoliths with focused 3B-27B health models. Same expert precision, 1/10th the compute footprint.

Local First

Edge Inference

Vision parsing runs on lightweight edge nodes (e.g., vLLM with Qwen 3B AWQ). The heavy 27B models wake only for complex metabolic cross-verification.

Zero Waste

Incremental Compute

We cache nutrition embeddings. When your intake doesn't fundamentally change, the AI doesn't re-compute. BMR/TDEE math runs purely on-device with zero network energy.

Roadmap

From private build to App Store.

Delemma began coding in November 2025. The app is in final integration; the nutrition language model continues training in parallel.

Delemma App for iOS 85%

Started Nov 2025. 85% complete — UI polish, biometric pipeline, and edge-case testing remain.

Delemma Nutrition LM 14B 60%

Domain-specific reasoning model on a separate branch — 60% trained. Ships at App Store launch.

Delemma World Food LM 3B 100%

Vision food-recognition model — trained, shipped and powering the live 49-nutrient pipeline today.

Nov 2025 Project kickoff

Apr 27, 2026 Final integration · 85%

May 10, 2026 TestFlight upload

May 20 – Jun 10, 2026 App Store release

Product

In your pocket.

Four screens, one loop. Log a meal, see the gap, get the move, close the gap.

Daily home dashboard — 01 · HOME Daily home Today's calorie balance, meal log and supplement check on a single glance.

02 · STACK Supplement intelligence Your stack, scanned, evaluated against today's gaps and goal — keep, swap or drop.

One-tap food recognition — 03 · CAPTURE One-tap recognition Photograph the plate. Get the full breakdown — calories, macros, micros, GI.

04 · LEDGER 49-nutrient ledger Every nutrient, percent of target, delta to close. Sorted by what matters today.

From the early users

It clicks fast.

A handful of testers running the loop for a few months. Different bodies, different goals, same pattern.

Sarah Mitchell

Distance runner · 34

Daniel Klein

Software engineer · 16:8 IF · 29

Akira Tanaka

Tokyo · cut 12 kg, kept it · 41

Maya Reyes

Fitness instructor · mom of two · 38

Ben Lawson

Biohacker · CGM + Apple Watch · 45

Liu Wei

Beijing · product manager · 33

The name

Why "Delemma"?

Every meal looks like a dilemma. Add the prefix and it stops being one — the way you de-bug code, you de-lemma a decision.

de-

Remove the burden. As in de-bug, de-fog, de-fuse — undo the thing in front of you.

lemma

A useful step. In math, the lemma you prove on the way to the theorem. Each data point is one — read enough, the answer reveals itself.

.ai

Where the agent lives. A homophone for Llama, a nod to the open-model lineage we build on. Intelligence as default, not feature.

Mission

Nutrition is a daily decision. Make it disappear.

Most health apps stop at calorie counting. The rest demand you read the literature. Delemma does the reading, runs the loop, and hands back a single next move — backed by the real numbers from your day.

From dilemma to Delemma.

Apply for external trial →