ADVENT
Field trials

Every result stays scoped to the test that produced it.

A sensor result means little without its conditions — what the target was, what else was in the air, how the trial was run. So we publish each test with its scope and never roll it up into a headline accuracy figure. These are ADVENT’s own verification results, not accredited third-party certificates.

Illustration of field verification with GraphAISense instruments
In the field

Data that exists only because an instrument was there.

Odour data cannot be scraped. Every trial below produced data that would not exist if an instrument had not been placed in a real location.

1,000+

Field trials run with GraphAISense instruments — the raw material every Model Pack is trained from.

3 partners

Public-sector field trials with Korea Environment Corporation, Gyeonggi-do and Incheon Port Authority.

19 units

Instruments deployed in public trials: 16 across Suwon, Hwaseong and Icheon, and 3 at Incheon Port.

January 2026 · in-house test

Smouldering wire insulation, against three background chemicals.

An electrical fault announces itself chemically before smoke or heat. The question was not whether it is detectable, but whether the instrument can tell the precursor apart from the ordinary chemistry of an industrial room.

TARGETABCPROJECTION 1PROJECTION 2
Target — smouldering wire insulationBackground ABackground BBackground C
※ A concept diagram explaining what separation means. It is not measured data, and neither the axis units nor the distance between clusters represents a performance figure.
Target
Overheated, smouldering wire insulation
Background present
Chlorine, ethanol and isopropyl alcohol — substances a single gas sensor cannot separate from a real precursor
Result
The target was correctly identified across the test set
What it does not say
Nothing about substances outside the set, and no general accuracy figure for the instrument
GraphAISense is a complementary monitoring channel. It does not replace statutory fire protection systems.
Q4 2025 · blind test

Separating odour classes, and resolving a mixture.

Real sites rarely emit one odour at a time. Before a reading can be acted on, classes have to separate reliably — including when they arrive mixed.

Diagram of odour classes separating in feature space
Classes
Chemical, wastewater and sewage odour classes, presented blind
Result
Classes separated with high accuracy; a mixed sample resolved into its two components
Why the mixture matters
Separating pure classes is the easy half — a reading that collapses under mixture is not usable in the field
What it does not say
The result applies to these classes under these conditions; it is not an overall accuracy claim
Field trials with public-sector partners

Where the instruments are working.

The organisations below opened their sites and helped set the test conditions. They are collaborators, not customers.

Illustration of field trial sites with public-sector partners
Odour source tracking with Korea Environment Corporation
Korea Environment Corporation · Oct 2025

Odour source tracking for residential complaints

Trained on four source odours and two mixtures, then traced three unlabelled samples back to their sources. Signed a performance-sharing agreement for environmental applications.

SEO Robotics patrol robot with e-nose at Incheon Port
Incheon Port Authority · 2026, in progress

E-nose on a patrol robot and ESS

Three units: an e-nose built into a SEO Robotics patrol robot in the port passenger terminal, plus fixed units at energy-storage facilities, linked to monitoring over MQTT.

Gyeonggi-do field trial installation
Gyeonggi-do · 2026, in progress

On-device AI fire detection

Sixteen units across Suwon, Hwaseong and Icheon — industrial complexes, traditional markets and public car parks — combined with AI CCTV in outdoor enclosures.

Bio & medical · research stage

From stool detection in critical care to breath testing

We are preparing stool detection for critical-care settings such as ICUs, where recognising by odour that a patient has soiled lets nurses respond sooner. The next step is breath testing — analysing volatile compounds in exhaled air. This is research stage; GraphAISense is not a medical device or a diagnostic product.

Partners are named only to record where trials run. Naming does not imply endorsement, and these are not purchase or delivery records.

How we report

Three rules we hold ourselves to.

We would rather publish fewer numbers than numbers that cannot be checked.

Rule 01

Conditions travel with the number

A figure without its target, interferents and trial design is not a result. If we cannot publish the conditions, we do not publish the number.

Rule 02

No generalisation

One scenario does not become an instrument-wide claim. Our 98% figure is per-scenario verification, not a blanket rate.

Rule 03

Self-verified is labelled as such

These are our own tests. Accredited testing (KTL, KOLAS-certified) is in progress; when a certificate exists we will name the body.

GraphAISense does not replace statutory fire protection systems, legally mandated measuring instruments or a battery management system. Healthcare and care applications are at research stage; it is not a diagnostic product.