Evidence before spectacle
Claims should be inspectable. Benchmarks include protocols and caveats; research tools keep the path back to sources visible.
Research · Applied and open
Question → protocol
Protocol → evidence
Evidence → product
Research is a product constraint.
At Amayx, research is not a decorative claim added after launch. It determines what can be said, what must be tested, and which limitations belong in the interface.
Current release · CalorieK
One photograph.
Five estimates.
Visible uncertainty.
CalorieK is an open 93.7M-parameter vision model that predicts calories, mass, protein, carbohydrate, and fat from a single RGB food image.
56.3 kcalmean absolute error on 606 valid locked dishes
30.0 kcalmedian absolute error
84.3%predictions within ±100 kcal
0 / 606invalid predictions
Protocol
Locked before the claim.
Every test image is a deterministic central frame selected by official dish ID from the Nutrition5k test split. Evaluation is dish-level, and the model card names the two physically impossible source labels excluded from the primary result.
Comparison
Same dishes, same question.
Published comparisons run competing checkpoints on the same inputs and report bias, invalid outputs, hit rates, and out-of-distribution behavior—not only a favorable headline number.
Release
Weights and limits together.
The checkpoint, usage code, training sources, benchmark artifacts, and explicit limitations are public. The model is research software and is not a medical device.
Method
From claim
to interface.
- 01
Define the decision
Start with the real user decision and the consequence of being wrong. This determines the evidence bar, not the novelty of the model.
- 02
Freeze the evaluation
Separate development from evaluation, audit leakage, record exclusions, and compare on identical inputs before choosing a headline.
- 03
Expose the boundary
Translate uncertainty and scope into product language: review states, source labels, local versus remote processing, and warnings where they matter.
- 04
Ship a reproducible path
Publish enough detail for another person to inspect, run, or contest the result. Reproducibility is part of the product surface.
Research across the portfolio
The same discipline,
applied differently.
Private by architecture
When a task can happen locally, it should. Network use is explicit, bounded, and explained rather than hidden behind convenience.
The user remains the author
AI proposes, compares, extracts, or drafts. People review consequential decisions, and exported work remains portable.