LuminaTechnology
What the sensor reads, 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 tell a contaminated surface from a clean one. In the lab, the research instrument resolves species. The field instrument is built to read bioload at the sites your program already swabs.
Every surface in a plant is a data source. Almost none of them are being read.
Signal path
From reflected light to a risk signature.
Units are being built to sit above drains, transfer points, belt returns and floor-wall junctions, the harborage sites your environmental monitoring program already samples, read continuously instead of swabbed weekly. Like a pre-collision warning in a car: it warns in real time, and the person still decides. 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 reference library built in the lab.
- 05 <10 ms
Flag
Above threshold, the read is written and an issue is filed for a person. Never a hold.
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.
Provenance
University research, licensed and acquired rather than reinvented.
-
Origin
An exclusive license to a seven-family university patent portfolio in hyperspectral bacterial detection, three patents issued, together with the acquired assets of the team that built the field instrument. For three of the seven families the license also conveys know-how, training and validation data, and software.
-
Core result
On cleaned stainless steel inoculated at stated loads, detection of E. coli ran 92.1–98.1% across six tasks and held above 95% down to about 40 CFU/cm², the lowest load tested so far. On spinach, 98.1% pooled and 89.1% on the hardest orientation. Lab results on our own instrument; independent validation is the work ahead.
-
Team
Engineers with backgrounds in remote sensing, machine vision, and data analysis. Our founder spent thirteen years building sensor networks at scale, nine at Flock Safety, where he was the first employee and became VP of Engineering; the network he architected passed 150,000 devices.
-
Data handling
Lumina runs in the cloud, with each customer's data isolated in its own tenant. Your documents are used to answer your questions and cite the passage; they are not used to train third-party models, and the assistant is not a shared wrapper over a general chatbot. We intend to train Lumina's own models on de-identified data, under the terms of your agreement, and never anyone else's.
Where it fits
A force multiplier for your program, not a replacement for confirmatory testing.
What it is built to add
- Which surface deserves the next swab.
- Which part of the line started drifting overnight.
- More of the plant covered, more often, on the same headcount and test budget.
What it does not replace
- The confirmatory laboratory testing your regulatory framework requires.
- Your release decision. A sensor flag is never a release criterion.
- Product and finished-product verification, on the schedule your HACCP plan and sanitation program already set.
In the lab, on inoculated stainless steel, detection held down to about 40 CFU/cm², the lowest load tested so far: the load of a surface where sanitation failed, not the trace levels an enrichment swab finds.
What comes next
Building the evidence base a category this new is going to require.
The figures above are headlines, 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 conditions, how ground truth was established, and the sensitivity, specificity and limit of detection behind the accuracy figures. Biofilms are untested and are in the validation plan. We will send what we have and tell you what we do not have yet.
Independent method validation and peer-reviewed publication are the work ahead. 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 are university-lab results on inoculated surfaces 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.
Regulatory
The rules changed. The paperwork got heavier.
Since January 2025, every FSIS sample from a ready-to-eat plant, product, food-contact or environmental, is reported for all Listeria species, not just L. monocytogenes. Any Listeria found obliges the plant to correct its sanitation. From early 2027, inspectors swab your surfaces with every routine product sample.
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.