LuminaTechnology
What the sensor sees, and how it decides.
Overview
Every material reflects light in a pattern set by its chemistry. A normal camera samples that pattern in three bands: red, green, blue. A hyperspectral camera reads hundreds of narrow bands instead, and at that resolution the pattern becomes specific enough to name what is on a surface.
Every surface in a plant is a data source. Almost none of them are being read.
01 · Signal path
From reflected light to a named organism.
Units sit above drains, transfer points, belt returns and floor-wall junctions, the harborage sites your environmental monitoring program already samples, watched continuously instead of weekly. What happens on each read:
- 01 0 ms
Illuminate
A fixed light source washes the surface at a known intensity and angle.
- 02 <1 ms
Split
Reflected light is separated into hundreds of narrow wavelength bands.
- 03 ~2 ms
Fingerprint
The band pattern is reduced to a signature: a curve, not a photo.
- 04 ~5 ms
Match
The signature is compared against the pathogen-linked reference library.
- 05 <10 ms
Flag
Above threshold, the event is written and the alert leaves the sensor.
02 · How the model sees a strain
See which cluster a read joins, and when it joins none of them.
A spectral read is not matched against one reference and declared a hit. Each sample lands in a space organized by strain, where what identifies it is the company it keeps, and where a sample that lands nowhere near a known cluster is flagged as unknown rather than forced into the closest label.
How the model sees a strain
Drag to rotate, or jump to another organism.
-
One read at a time
Every dot is a single spectral reading taken off a surface. On their own, they are just numbers.
-
Structure appears
Group the readings by similarity and three regions form, one per organism. Nothing labeled them. The readings simply resemble each other.
-
Inside one organism
Zoom into E. coli and the pattern repeats. Each strain we know keeps its own tight neighborhood.
-
The same holds everywhere
Salmonella and Listeria behave the same way. Organisms sit apart, and strains sit apart inside them.
-
A new read arrives
It lands inside a neighborhood we already know, so it can be named, matched to a strain the system has seen before.
-
A read with no neighbors
This one lands in open space. Something is there, and it is nothing we have characterized, so it is flagged unknown instead of forced into the nearest label.
-
The whole picture
Every read the system has seen, in one view: known strains sitting in their own neighborhoods, and the unfamiliar one out on its own. That separation is what lets the system alert on known pathogen signatures while still catching unknown variants.
Illustrative: cluster geometry is simulated to explain the approach, not rendered from a trained model.
03 · Provenance
Thirty years of imaging research, acquired rather than reinvented.
-
Origin
A portfolio of sensing and analysis IP: roughly 30 years of hyperspectral imaging research, with the last five-plus years focused specifically on bacterial detection and classification.
-
Core result
In 2024, that research demonstrated 95% accuracy detecting E. coli across surfaces spanning product (spinach and chicken breast) and equipment: stainless steel and plastics.
-
Team
Engineers with backgrounds in remote sensing, machine vision, and data analysis. Our founding team spent fifteen years building imaging and evidence systems for public safety.
-
Data handling
We build and run our own models in our own environments. Your documents are not sent to a third-party AI API.
04 · Where it fits
A force multiplier for your program, not a replacement for confirmatory testing.
What continuous sensing multiplies. Your team already knows how to investigate a positive. What it does not have is a way to know which surface deserves the next swab, or which zone started drifting overnight. Continuous sensing points them at it, so the same headcount and the same test budget cover more of the plant, more often.
Where the lab still rules. It does not replace the confirmatory laboratory testing your regulatory framework requires, and it is not a release criterion. Your product and finished-product verification keeps running exactly as it does now, on the schedule your HACCP plan and sanitation program already set.
05 · What comes next
Building the evidence base a category this new is going to require.
The 95% figure above is a headline, and headlines are not how method performance gets evaluated. If you are assessing Lumina seriously, ask us for the underlying study detail: organisms and strains, surface matrices, sample counts, controls, soil and biofilm conditions, how ground truth was established, and the sensitivity, specificity and limit of detection behind the accuracy number. We will send what we have and tell you plainly what we do not have yet.
Independent method validation, peer-reviewed publication, and a named scientific advisory board are all in motion. We are building that record deliberately and in the open, because a company that intends to change how an industry proves food safety has to be able to prove its own.
Accuracy figures on this page refer to the acquired research program as of 2024 and are not a claim of independent third-party method validation. Performance in a specific plant depends on surface, soil, sanitizer residue, and installation conditions.
06 · Regulatory
The rules changed. The paperwork got heavier.
Since January 2025, FSIS tests for all Listeria species, not just L. monocytogenes, across product, food-contact, and environmental surfaces. A positive is treated as evidence the sanitation program is not preventing the conditions where Lm takes hold, and it carries a corrective-action obligation.
What changed and when, with the primary FSIS sources. What a single positive now costs you. Which zones the rule reaches. And where continuous sensing fits against the confirmatory testing the rule still requires.