WORKING WITH US

Step 1: Your data

Do you have data that captures the key variables of your process? Most organisations are drowning in data, but a common worry is that data is often spread across different systems or of questionable quality. AI needs data and data feeds and so this is always the first consideration once the drivers for the AI solutions have been established. Pulling from variable data sources, verifying data, and cleaning data sets of errors and misleading readings is a standard part of this process.

Step 2: Selecting or creating a data model to fit your application

A model will describe how the variables interact. For example, a bread making data model might describe how baking time changes with different oven temperatures, ingredient quantities and baking moisture levels. Sometimes research has already been done to investigate the relationship between the variables and a data model can be selected from online open sources. However, in many cases a new model will need to be created by examining the data to confirm various hypotheses e.g. will a lower oven temperature and longer cooking time result in a loaf of better or inferior quality? By examining how multiple variables interact, and seeking statistically reliable relationships between them, a best-fit data model (sometimes named an algorithm) is created that can describe expected outcomes from given inputs. Complex AI solutions often involve multiple data models.

Step 3: Teach and test

This is an important step where the full strength of the data is converted into ‘artificial experience.’ Teaching can be both a human process and a machine process, typically with a process expert providing input on some of the significant real events that might have happened and describing how it shows in the data. The machine is left to crunch the years of more mundane data.

The model is tested for accuracy by allowing it to predict results and then observing how closely they match real world outcomes. Final calibration and adjustments can be made to improve the match.

Step 4: Operationalise

Arguably, this is the step that differentiates AI research projects from commercially deployed AI solutions, for example Alexa having immediate answers and Tesla in driving mode without crashing. These AI models work because they are almost perfectly operationalised i.e. the theory is matched by the way they work in reality. Not all solutions need to operate in real-time, but all solutions need to have robust data feeds and an ability to compute quickly so that the user gets feedback as soon as it is needed. Scalability to deal with large quantities of incoming data or users is also an important element of operationalisation. Typically the hardware for measuring and feeding latest data into the data model needs to be in place to support the speed of the process. Cloud computing is often the facilitator in many solutions allowing input data to be computed in microseconds and served back to the user with an answer. Whatever the AI challenge, how to operationalise the solution so that the user’s experience is seamless is a critical step.