➡️ When working on an intelligent elevator we looked at its hardware and tested our solution ourselves live.

➡️ Working on a personalized medicine project, we visited a hospital to understand how new patients move from an emergency room to a hospital ward, and what data is collected in this process and how.

➡️ We visited a sawmill while discussing algorithms for cutting logs optimally.

➡️ We counted money when working for a bank. Just kidding this time.

Why are we doing this? To understand what a process looks like that we want to improve and how exactly the results of our work will be used.
Despite (or thanks to) well over a hundred projects under our belt, we know we don’t know everything.

AI/ML projects are very sensitive to details. We call them “fragile”. They are full of risks.

According to Gartner’s report, 85% of AI/ML projects underperform due to the quality of the data, algorithms or teams responsible for implementing them.
Contrary to Gartner’s reported figures, 95% of our projects are successful. For the remaining 5%, we quickly recognise when data does not adequately describe the problem. In such cases, we promptly pivot or stop the project, thus avoiding unnecessary costs.

By the way, 80% of our projects go into production.

We achieved this success ratio thanks to our proprietary methodology for executing ML/AI projects – Quantup Thinking.
It’s based on three pillars:

🟢 comprehensive knowledge of the project business domain,
🟢 deep understanding of the applied ML/AI methods,
🟢 a highly standardized software development process.

Deep knowledge of the business domain, including industry-specific language, is key to defining project goals and translating them into mathematical models. It also ensures clear communication throughout the project, especially when discussing data or the impact of modeling decisions or process changes.

A deep understanding of the methods used is the foundation for making informed choices. Each method has its limitations, complexity, and level of explainability of the generated results. They are significant from the perspective of its potential application. The last aspect can be a decisive criterion for method selection, especially in the context of mass decision-making, due to the business risks involved.

Our coding environment provides:

🟠 support for good design patterns,
🟠 support for ensuring appropriate quality,
🟠 automation of containerization and computation processes;

It’s an environment for developing successive model versions, which means it ensures:

🔴 reproducibility of computations,
🔴 comparability of results obtained for different model versions.

Quantup Thinking is inspired by:

🔵 IBM CRISP-DM,
🔵 Lean management,
🔵 Agile approaches,

and focuses on:

🟣 measurability of outcomes,
🟣 repeatability of experiments,
🟣 an iterative approach for incorporating increasingly larger datasets and more complex modeling techniques,
🟣 effective solution integration (training vs. production data).