limel

For the type of application you describe, we are planning to incorporate nanobio detectors, which can be extremely sensitive. In some applications we already combine non-selective and selective sensors, and together they create a multidimensional digital signature that is analyzed by our machine learning models.

If you have a specific problem in mind, please feel free to reach out to us via the Sniphi website. We would be happy to explore whether this is something we could support you with.

Thanks for sharing these insights and real-life examples, they are very interesting. Yes, I believe there are situations where an olfactory sensor could detect a problem earlier than conventional sensors. For example, the smell of burnt oil might appear before an oil level sensor detects a drop or before temperature sensors trigger an alarm.

The key question, however, is where the biggest value lies - either in cost prevention or hazard reduction - so that the benefit-to-cost ratio justifies the investment in a technology that is not traditionally used for this purpose.

What you mentioned is particularly important: identifying industries where people already rely heavily on the perception of odor. Not just selective measurement of a specific chemical compound, but the overall “human” impression of smell. In many environments an experienced worker with 20 years in the industry can simply smell that something is wrong.

If we can replicate that capability with Sniphi - but in a scalable, continuous way - it could make the value proposition much easier to demonstrate to customers. Thanks for sharing.

You nailed it with the ‘poocalypse’. My parents have two cats, and that exact scenario was the reasons they decided not to buy a cleaning robot.

Technically, I’m quite confident that Sniphi could recognize the smell of poo. The bigger challenge would be the environment inside a cleaning robot — dust and particles could interfere with the sensors. I believe this could be addressed with some engineering effort, for example using filtration or protective sensor covers, but it would require additional work.

That’s why we are also looking at some lower-hanging applications. For example, in nursing homes or hospitals. A bedridden patient with a diaper that isn’t changed in time can develop serious complications — especially elderly patients, where infections can become life-threatening. In that context, a “sneaky poo detector” could actually improve care and potentially save lives. Do you know anyone who might be interested in partnering with us to test this idea?

Azure, together with Power Platform tools such as Power Apps, is primarily used for large-scale training. Since the odor and gas data require efficient labeling to be reliable, Power Apps combined with Data Explorer provides an easy, cost-effective, and scalable way to manage this process. Once trained, the model can be deployed directly on the edge.

Yes, many current solutions rely on VOC sensors such as those from Sensirion, and we share the observation that a lot of the existing implementations remain at the demo stage and struggle in real industrial environments.

In our case we are somewhere in between research and real deployment. We recently completed research on detecting food infestation — across different insect species and different food samples, and the results were successful enough that we submitted the paper for publication.

At the same time, we are fully aware that this is not yet an industry-grade solution tested in environments like grain silos. For example, heavy dust is a major factor that can affect sensor performance without proper filtration, so there is still engineering work to be done before large-scale deployments.

However, we believe the core detection capability is already promising, and we are now focusing on solving the practical challenges needed to move from research results to robust industrial applications.

lol, that would be fun

That’s a very good point, and we actually see fermentation batches as one of the most promising early use cases. In many facilities, smell is already used informally as a pass/fail indicator, but it is subjective and difficult to scale.

We measure both humidity and temperature and use them as additional inputs for the ML models. Regarding sensor drift, it is still difficult to fully assess its impact on the business case. At this stage, our main focus is on the accuracy of the classification models rather than very long-term operation — that would be the next step.

For now, the practical approaches we consider are either on-the-fly calibration through a feedback loop based on the actual process output, or simply replacing the sensor when necessary, as the manufacturing cost is relatively low.”

That’s true. We even started a PoC with a skincare products factory. The challenge, however, was that the frequent rotation of the product portfolio — and the large number of SKUs — made it difficult to justify the training effort.

On the limits of detection - with Sniphi we follow a different approach than traditional selective sensors. The system is based primarily on non-selective chemical sensors operating at controlled temperature profiles. Each measurement cycle (6 seconds) generates around 60 measurement points per sensor, creating multidimensional signatures of gas mixtures that are then analyzed using classification models.

That is actually an interesting direction. Since smoke detectors already exist, the next level could be distinguishing smoke from a cigarette — or even something harmless like burning scrambled eggs — from more dangerous sources such as burning carpet or electrical wire insulation. We will definitely think about it.

A mold detector is also an interesting idea. Our ‘digital nose’ can measure humidity and temperature as well, and these factors are often strongly correlated with mold growth. Combining odor detection with environmental data could therefore be very useful for early mold detection.

Unfortunately, medical applications require enormous time and effort to meet strict verification and regulatory requirements. While this is an important long-term direction, we are currently focusing on lower-hanging opportunities such as food manufacturing and processing, where there is strong potential for cost savings and loss prevention.

Digital smellers are scalable and more repeatable than human noses. At the current stage our electronic nose operate either through classification of previously trained odor classes or through anomaly detection. What is still missing is a possibility to run a more sophisticated conversation with the model when something smells "suspicious".

The problem is not whether we can digitize the sense of smell, but that no industrial process currently relies on it by default. The real challenge is identifying the first scalable use case that proves measurable business value (sniphi team member here).