5 days ago
Artificial Intelligence (AI) is transforming industries across the globe, driving efficiency, reducing costs, and opening up new opportunities for growth. However, one of the most impactful aspects of AI lies in its ability to be customized to solve specific, unique challenges that businesses face.
In this article, we explore several real-world examples of how custom AI solutions are being utilised to address unique business problems.
How good are standard models?
AI is often associated with LLMs (like ChatGPTor Gemini), or with generative models (like Dall-e, Midjourney) that give complete freedom to your creativity. As the name suggest, Large Language Models are large and therefore require a lot of computing capacities and energy. Also, at first glance, they might look like they are good at everything. However, ensuring the required level of accuracy and dealing with hallucinations might be difficult to deal with.
Taking a ready-to-go ML model and then deploying it in the production may seem like the easiest thing to do, but there are many factors to consider. The challenges may include data privacy, the specific nature of the problem, processing time, scalability, explainability and many others.
Another important thing to consider is the legal situation. First of all, some models or even the datasets that they were trained on may have closed licences so using them might not be possible.
Many businesses focus on solving a specific problem. This can mean that off-the-shelf models do not perform well in this particular case. Also, the issue might be that the data that is not easily accessible. Custom solutions take a more creative approach that opens more doors. With this approach, data scientists can combine their experience with the support of the client, who provides the domain knowledge.
To sum up, there are a few things to consider:
Benefits of Custom AI Solutions
- Tailored Fit: Addresses specific business needs and challenges.
- Scalability: Custom AI solutions grow with the business.
- Data Utilization: Leverages proprietary data to gain a competitive edge.
- Improved ROI: Provides targeted outcomes and measurable results.
Challenges and considerations for implementing AI
Although custom AI offers significant benefits, businesses must consider factors such as:
- Data Privacy: Ensuring compliance with data protection regulations.
- Cost: Higher up-front investment costs compared to off-the-shelf solutions.
- Integration: The need to integrate AI seamlessly into existing systems.
- Talent: The need for skilled data scientists and AI engineers.
Some of our custom solutions
Defects in a toy factory
There is huge potential for custom ML models in manufacturing. For example, the visual detection of defects saves a lot of time and money because a machine can work tirelessly and much faster than a human. Also, some differences might be very hard or even impossible for the human eye to spot. We worked on such a problem in a toy factory by detecting printing errors on bricks.
Automation with the use of very simple convolutional neural networks allowed us to identify shifts in the printout and spots of ink. The output from the network was the probability of an error in each pixel in the form of a heatmap.
The models we used were tiny, especially in comparison to the networks used in the industry. This allowed us to run the model on an edge device with low inference time.
Railway anomalies project
An obstacle for regular models might be the type of input data. Most of vision models operate in the visible light spectrum in RGB format. Some phenomena can be understood much better in different modalities. One of the most popular examples is thermal imaging. In this domain, finding the data is even more challenging than in “regular” domains. The computer vision-related problem that we were solving was the detection of anomalies in the infrared spectrum, where overheated devices were much easier to distinguish.
An example of the images that we were dealing with are presented below. In the picture, you see a blue box, which represents an anomaly and a red box for detection.
In this case, the best approach was dividing the image into tiles, then using an object detector that was fine-tuned on the data that we provided. The challenge here was to reduce the number of false positives, the elements where higher temperature does not necessarily mean that something is wrong. The final solution selected only “suspicious” tiles and forwarded them for verification by a human.
Using this solution, we were able to reduce time and cost by 84%.
Target classification using radar data
Another modality with a great deal of potential to be used in machine learning models is radar data. Our client wanted to increase accuracy in the detection of predefined classes based on the data from FMCW radar. Radars use the Doppler effect to “see” objects in motion.
By using 1-dimensional convolutional neural networks, we were able to increase detection accuracy up to 98.44%
Predicting device reboots
Our client provided us with data from home routers so that we could predict when the performance of the routers would drop. In such a case, a reboot can be performed to bring back the initial state. This problem falls into the category of predictive maintenance. Neural networks were not needed in this case because the best-performing solution was XGBoost, an algorithm that is based on decision trees.
The solution was applied using data from over 10 million devices. The accuracy of the model was over 85% with a precision rate of over 80%.
Conclusion
As you see in the above examples, standard out-of-the-box projects are not sufficient for every task and some level of expertise is required to solve a specific problem. Our AI solutions integrate smoothly with existing systems, which helps to accelerate development and improve operations without disruption. This ensures that the solution brings measurable ROI, whether the client needs quick gains in efficiency or long-term growth in the business. When highly experienced data scientists combine their skills with the domain knowledge provided by the client, then truly amazing outcomes can be achieved!