How to set data and infrastructure for an AI solution. Rules of Thumbs.
Working in AI implementations for many different industries and projects, I have been exposed to various different vendors, infrastructures and data architectures.
The variety of purposes and applications is so large that I can definitely say there is not such a thing like a ‘best AI infrastructure approach’.
I can say though there are definitely the same mistakes and misconceptions commonly occurring no matter which industry is adopting or implementing an AI model.
Here the most common:
- Implementing an AI model without any cost plan for its infrastructure. In terms of line of codes, actually integration software, APIs etc are 95% of the whole live AI solution. Only 5% of the code is specifically the shiny AI model. Hence the actual cost, over the long term, is on the infrastructure.
- Data maintenance and preprocessing may be an ongoing cost, not a setup cost. Since the data science cleaned the data and we have verified the AI solution performs as expected, there is the temptation to disregard further data checking and preprocessing once the solution is live. In the end, what can go wrong? If there is a bug, as any software, we are going to fix it when it occurs. Bad mistake. An AI solution is not a ‘static’ software and re-processing a large dataset and pipeline after things have gone wrong is not a just matter of bug fixing. The model at that stage probably has been training with wrong data and you do not know exactly when. Good/bad data are likely mixed. It is way more complex than just fixing the bug and the system resumes to normal.
- Cybersecurity and MLOps are not cost effective from the start. Sometimes great AI solutions with great performance metrics sadly cannot go live because the machine learning operations are simply too complex and too costly (too much need of retraining the models, too frequent data checking, versioning etc. MLOps). Same goes for costs attached with security.
#ai #artificialintelligence #business #innovation