Building an AI Process Engineer

Most of the recent news about artificial intelligence has been about foundational technology. You hear of new large language models, the latest GPU, and scads of data centers being built all over the world. Which is great for AI geeks of course; we need all that stuff to make AI solutions run.?

The true promise of AI, though, does not lie in all this foundational technology, but in applying all that amazing tech to the task of solving complex problems, which has been the sole purview of good old fashioned humans for eons. We’re just beginning to learn how to use AI to truly tackle the messy, intricate, stubborn challenges of the real world.??

This is what we’re focused on at TC Labs: not inventing new AI models, designing faster chips, or building cloud platforms, but developing tools and systems that allow businesses to apply that raw technology to specific, complex situations. In our case, those complex situations are the ones that plant managers in refineries and other factories see every day.?

To create a system that solves problems like humans do, it helps to ask the question “how do humans solve problems?” By our thinking, there are three relevant steps. First is planning. Our human comes up with an approach and a plan to solve the problem, executes the plan, then learns and tries again if it fails.??

Second is tools. Our human figures out which tools they need to carry out the plan, and if they don’t have those tools on hand, procures the ones that are missing.??

Third is data. Our human gathers the data necessary to fully analyze and solve the problem, and knows which data is the most important to get it right. They understand the semantics of the data, which vary widely from plant to plant, and discern which data matters the most for this particular situation.?

There's another critical ingredient: expertise. A great plan and a full toolkit likely won’t go far in the hands of a novice. Our human gains that expertise through education (e.g. school) to build foundational knowledge, and by learning the details of the particular situation (e.g. a refinery). Then they get smarter through doing. Everytime they run across something new, they learn a bit more and naturally file it away for future reference.

It takes time to build expertise, which is why there are never enough experts, they are always busy, and they can be quite expensive.?

We mimic this problem solving approach with our TC Labs AI process engineer for refineries and other types of heavy industry factories. When presented with a real-world problem, our platform creates a plan to solve the problem, identifies and procures the tools it needs to execute on the plan, and pulls necessary data from a wide array of sensors and data stores, prioritizing the most important metrics based on the identified problem. It iterates on its plan based on internal scenario modelling and then real-world feedback, honing in on the best possible solution for today’s problem while learning for future scenarios.

To make this system an expert system, we arm it with large language models that provide an intellectual foundation. We educate it with detailed knowledge and data about a particular plant: logs, schematics, manuals, emails and presentations, interviews with staff, and even handwritten notes. (This education usually can be done in a few days.) Finally, we feed it a bunch of scenarios and guide it through an iterative learning process.?

When a complex problem arises in a plant (as it does almost everyday), the TC Labs platform is like a skilled and knowledgeable expert system on hand to help the staff, one that is well-versed in how to solve problems.?

Let's say a plant operator wants to know if a reactor is healthy? The TC Labs system starts by studying up on what are the problem signs of an unhealthy reactor, determines the metrics for those symptoms, retrieves data from the plant, and isolates the metrics most important for performance. It compares the real values with the expected ones to see if something is off, and pulls local logs to see if any problems have already been recorded. It synthesizes all this data across the complex system into a concise report and an answer to the question. All in just a few moments.

If the operator wants to investigate how to improve performance, say by changing the temperature in a furnace, our system creates a model that predicts the relationship between furnace temp and output. This model can use historical data, including times where adjustments didn’t work, and it can also integrate other approaches based on foundational knowledge. It scans the data from previous adjustments to see what worked before and what didn’t, runs its model to identify additional parameters that affect output, gathers current data on those parameters, and puts it all together to project the output change. Now the plant operator can try out different tactics and have a very good idea of which works best before making an actual change to operations.??

To improve our system’s problem solving capabilities, we constantly feed it new scenarios, see how it does, and guide it to the best strategy and plan. Much of this education is localized, based on the problems that a particular plant has or might experience. This step is similar to the process we employed when I was at Google to constantly improve our search results (as far as I know, they still use it). We employed people to feed the search engine challenging or new queries and rate how it did. These ratings were shared with the engineering team, who used them to tweak the algorithms. Our system, just like Google search, never stops learning.??

Every factory is different in its own idiosyncratic ways, so the ability to quickly make a system a constantly learning expert on that particular plant is critical to get to effective problem solving quickly. The nice thing is, AI is a very good student. You only need to teach it something once and it remembers it forever.

Think about the last time you went to see a primary care doctor. The physician has an impressive foundational education (med school), lots of hands-on training (internship and residency), and a bunch of tools and specialists at their disposal. She applies all of that knowledge, skills, tools and specialists along with her unique knowledge of her patient (you!) to prescribe the best possible care plan.?

TC Labs is like that doctor, only for complex factories such as refineries. We take all that amazing AI technology and build a bridge to solving everyday problems for plant operators.??