Vocify’s Electronic Nose Achieves 98% Real-World Accuracy with Google Cloud

About Vocify

Vocify is an innovative startup that has developed a sophisticated “electronic nose,” a device capable of detecting and distinguishing a wide array of odours. Their mission is to replace traditional, often inefficient, manual methods of scent detection. Driven by a vision to transform how the world uses smell to solve real-world problems, they are pioneering the integration of scent technology into healthcare, agriculture, security, and environmental monitoring.

The Challenge

Vocify faced two major, sequential challenges regarding their electronic nose data and models.

Initially, their manually built AI model failed in practical applications despite achieving 99% accuracy in the lab. The model suffered from overfitting (memorising training data) and data leakage, which caused its accuracy to collapse in unpredictable real-world environments. This stalled growth and jeopardised key customer deals.

Following initial improvements, the core challenge remained: how to interpret, analyse, and build high-quality models consistently. Vocify needed to move from using offline, manual prediction methods to providing live, on-site detections. When the customer attempted to implement their manually trained models in live, real-world scenarios, a huge performance drop was noted, and the manual model generation process was slow and laborious. key business opportunity. Our goal was to build a model that was not only theoretically accurate but also practically robust.

The Goal

The ultimate goal was twofold: 

• to build a practically robust and commercially viable machine learning model that could reliably generalise to unseen data
• to establish an automated pipeline capable of delivering live, highly accurate predictions on site. They needed a solution that would deliver superior quality in a fraction of the time, transforming their prototype into a reliable product.

The Solution

Devoteam delivered a comprehensive, two-stage solution using Google Cloud technologies, fundamentally rebuilding Vocify’s data and ML pipeline. The experts first rebuilt the model for better performance and reliability, and then worked on improving its speed and addressing the slow manual generation process with VertexAI AutoML.

The Methodology

Devoteam delivered a comprehensive, two-stage solution using Google Cloud technologies, fundamentally rebuilding Vocify’s data and ML pipeline.

Stage 1: Rebuilding for Reliability

Recognising the symptoms of an overfitted model, Devoteam rebuilt the machine learning model from scratch. This began with meticulous data cleansing to remove redundant or irregular data, ensuring that the training sets represented real-world conditions. The new approach centred on generalisation, sophisticated feature engineering, and feature selection.

Devoteam initiated the work in Google Colab, automating data pipelines and establishing a robust testing framework as a pragmatic “stepping stone” towards a full production environment.

Stage 2: Automating for Quality and Speed

To achieve live predictions, Devoteam set up Google Cloud Storage to host models and Google Cloud Run to communicate with them, facilitating the switch from manual offline predictions to live detections.

To address the performance drop noticed in real-world testing and the slow manual generation process, Devoteam advised the customer to adopt Google Vertex AI AutoML. This crucial implementation drastically reduced the time required for model generation and simultaneously increased model quality.

Results

Devoteam’s solution transformed the electronic nose from a prototype into a commercially viable, high-performance product, delivering measurable results across reliability, speed, and accuracy.

• Reliability and Commercial Impact: The re-engineered solution delivered a 2x improvement in accuracy on real-world samples, transforming the device’s reliability. This foundational success created a compelling new value proposition.
• Increased Speed and Efficiency: The implementation of Vertex AI AutoML drastically reduced the time and labour required for model generation. The model generation cycle was cut from approximately two months to just one to two weeks, speeding up product development and iteration cycles. The process is now performed almost fully automatically, eliminating manual labour.
• Enhanced Accuracy: The final solution significantly enhanced prediction accuracy in specific use cases. For detecting bedbugs, the system’s accuracy increased from around 90% to almost 98%, demonstrating a substantial, data-driven improvement in model quality.