Beyond the AI Hype: What’s Actually Working in Commercial Drones?

Now, that's a great question that also keeps me up at night all the time.

When you see, I'm in San Francisco, so everything is AI, all the billboards are AI, everything's agentic.

Every bus has ads on agentic workflows and it's kind of surreal to see that.

But in our industry, one of the key things is our core models are still deterministic, which is we are using our traditional computer vision models.

to find defects and the main reason for that is there's more guardrails there's more control and then there's explainability behind uh you know what is being detected so one of the key things in utilities are these are mission critical projects there's little room for error if there's error and there's hallucination in the models that means hey, a substation equipment can go down, a transformer can explode that might affect society in a negative way.

So we really rely still on our deterministic models.

At the same time, we still use stochastic models, which are our vision language models that you see, the generative AI models.

We use them to augment or make our deterministic models more powerful.

One example is we use generative AI models for generating synthetic data.

There's examples where we won't have a lot of real-world data, so we might have 20 pictures of an insulator that's flashed, and we want to generate 5,000 pictures of it to train our core deterministic models.

Sometimes we use VLMs, vision language models, to augment that, and that gives pretty good results on that side of things.

At the same time, we have identified certain things where there is more range for error and hallucination, even though we still put harnesses and guardrails around these agents as well.

One example is once our models detect what kind of defect is happening on the equipment, The next thing is the utility want to know is what kind of maintenance can they do? That's where a new agent that we have built uses guidelines from FERC and NERC that's in public domain, and they do recommendations that, hey, this is a priority zero emergency defect.

You need to get out in the field within 48 hours.

You need to do this in terms of steps, and that really helps our utility customers to drive outcomes.

So, at the end of the day, I would say for us, it's a mix of both deterministic and stochastic models.

You just have to know what the right use case is and how that could drive the business outcome.

Got it.

Just a quick follow-up.

So, if somebody is looking and evaluating what solution to use, right, how do they take that call, right? How much do they go to the new models or the older models or how do they decide as a utility company or a power company?

Yeah, I think that's an interesting question because there's multiple factors that affect it.

The first is I would say unit economics is a big thing.

You know, generative AI models, they're expensive.

Tokens, yeah.

No more token maxing for companies.

Now, that's a great question that also keeps me up at night all the time.

When you see, I'm in San Francisco, so everything is AI, all the billboards are AI, everything's agentic.

Every bus has ads on agentic workflows and it's kind of surreal to see that.

But in our industry, one of the key things is our core models are still deterministic, which is we are using our traditional computer vision models.

to find defects and the main reason for that is there's more guardrails there's more control and then there's explainability behind uh you know what is being detected so one of the key things in utilities are these are mission critical projects there's little room for error if there's error and there's hallucination in the models that means hey, a substation equipment can go down, a transformer can explode that might affect society in a negative way.

So we really rely still on our deterministic models.

At the same time, we still use stochastic models, which are our vision language models that you see, the generative AI models.

We use them to augment or make our deterministic models more powerful.

One example is we use generative AI models for generating synthetic data.

There's examples where we won't have a lot of real-world data, so we might have 20 pictures of an insulator that's flashed, and we want to generate 5,000 pictures of it to train our core deterministic models.

Sometimes we use VLMs, vision language models, to augment that, and that gives pretty good results on that side of things.

At the same time, we have identified certain things where there is more range for error and hallucination, even though we still put harnesses and guardrails around these agents as well.

One example is once our models detect what kind of defect is happening on the equipment, The next thing is the utility want to know is what kind of maintenance can they do? That's where a new agent that we have built uses guidelines from FERC and NERC that's in public domain, and they do recommendations that, hey, this is a priority zero emergency defect.

You need to get out in the field within 48 hours.

You need to do this in terms of steps, and that really helps our utility customers to drive outcomes.

So, at the end of the day, I would say for us, it's a mix of both deterministic and stochastic models.

You just have to know what the right use case is and how that could drive the business outcome.

Got it.

Just a quick follow-up.

So, if somebody is looking and evaluating what solution to use, right, how do they take that call, right? How much do they go to the new models or the older models or how do they decide as a utility company or a power company?

Yeah, I think that's an interesting question because there's multiple factors that affect it.

The first is I would say unit economics is a big thing.

You know, generative AI models, they're expensive.

Tokens, yeah.

No more token maxing for companies.