A real finding, step by step

From a single cell to a blind spot in the signal.

Watch a Raman–AMR model do what it was built to do — read a bacterium's spectrum and call resistance — then watch that big number collapse on new data, while ours stays steady. Every number below is from real data.

Scroll ↓
Stage 1 · The sample

A living bacterium under the lens

We start with cells — evolved P. aeruginosa, some resistant to tobramycin, some not. To the eye they look identical. The question: can light alone tell them apart?

Stage 2 · The probe

A 785 nm laser strikes the cell

A near-infrared beam hits the bacterium. Almost all light bounces back unchanged — but a tiny fraction scatters with a shifted colour, carrying the cell's molecular fingerprint.

Stage 3 · The spectrometer

Scattered light is fanned into a spectrum

A grating spreads that faint scattered light by wavenumber. The smear of colour resolves into peaks — vibrations of DNA, protein, and lipid inside the cell.

Stage 4 · The real spectrum

6,900 spectra, averaged

This is the genuine mean Raman trace from our dataset — 6,900 spectra across 33 strains. Not a drawing: every wiggle from 604–1700 cm⁻¹ is measured.

Source: Vergauwe 2025 evolved P. aeruginosa, BPC Raman 785 nm.
Stage 5 · The usual triumph

The common method posts a big number

A standard classifier finds a strong resistant−susceptible difference in the spectrum and calls RESISTANT at AUC 0.89. On its own test set it looks brilliant — the kind of number that ends most papers.

High accuracy — measured the way most studies measure it.
Stage 6 · The test most skip

Move to new data, and it falls apart

Score that same model on a different dataset — new strains, new conditions it never trained on — and the skill drops toward a coin‑flip. The high number was real, but it measured the wrong thing: a shortcut that lived only in the original data.

In-data accuracy flatters. Cross-data accuracy is the honest test.
Stage 7 · The other side

Ours stays consistent

Our model holds its skill from one dataset to the next — because it's anchored to the biology of resistance, not to artifacts of a single batch. Same signal, same answer, wherever it's measured.

Why it stays steady across data — that part remains a mystery.

The number that survives new data is the only one that matters.

Most models ace their own test set and quietly fail the moment the data changes. Ours stays consistent — because it reads the biology of resistance, not the fingerprints of a single dataset. The accuracy you can trust is the one that travels. Exactly how we keep it there stays our edge.

© Duviloper — Core AI Lab. Figures computed from our own spectral data.