Last year I had the pleasure of discussing what we do here at Inviol with talented journalist Mary Hurley from Caffeine. We covered topics like challenges with AI, AI trends, and what AI can offer startups. Below is a snippet of our chat. To read the full article, click here.
How do you use AI in your daily operations?
There’s some machine learning in there but it’s mostly computer vision and how to get the computer to see, estimate depth and do all those things properly.
We use a lot of open-source software and models. We reference architectures that others have developed and then tweak and refine those to build our own models.
A crucial part of the recipe is how much data you have. We have millions and millions of data points for training these computer vision models to recognise different objects and scenarios.
Then, we use a bunch of tooling to help our developers work faster. So Github Copilot for DevOps and the likes of ChatGPT to accelerate coding processes and things internally.
How are you fostering AI literacy internally?
It comes down to cultivating a culture around sharing and being intrigued by AI. We’re quite lucky as our team is really interested in the subject. We have a chat room just for new things people have seen that might or might not help us.
We find LinkedIn is a really good source as long as you follow the right people – technical people always talk about what they’re doing or where something is heading. Whenever anyone publishes something exciting, we’ll read over and discuss it.
Then, as not all of our team are building AI, we make sure to do demos as part of our processes. We’ll share what’s being built and how it works and get a good understanding amongst the whole team.
What has been your biggest challenge with AI?
A big challenge has been building generalised models that fit into multiple environments without requiring lots of retraining or having specific models for different customers.
That’s something we’ve been able to overcome, which we’re quite happy about.
How do you manage privacy concerns?
We align ourselves with New Zealand privacy laws.
From our end: one, there’s no one watching because it’s a system, and two, we don’t send anything anywhere unless it’s a health and safety breach or showing signs of unsafe behaviour.
We also integrate it with technology and frameworks that keep people’s information private when needed, be that facial blurring or data encryption. And, whenever we have trucks going onto customer sites, we make sure to notify people that a camera is there.
Does government legislation help or hinder you?
I know it’s starting in the US and other countries, but New Zealand doesn’t have much at the moment – what there is lines up with standard technology legislation.
So far, we’re playing it pretty safe. We’re not building large generative models and recreating people’s work. For us, it’s more about managing the privacy concerns I mentioned. As long as we stick with protecting people’s privacy rights, we’re good.
Are you monitoring any AI trends?
I’ve got my eye on Edge AI. It does the computation live on site. That means the data doesn't have to be shipped anywhere; it's not going up to the cloud to be processed and coming back down.
Most of the competition does the processing in the cloud, which is slow and can’t turn on alarm bells or send alerts fast enough. We have the hardware for doing edge processing installed in the trucks, which means we can do real-time learning.
It’s going to be huge; it’s already starting to be. Over the next three to five years, I think there’ll be a real shift from centralised cloud-based models to personalised Edge AI models on individual devices such as phones and laptops, creating tailored user experiences.
What does AI offer startups?
It allows us to be faster and do more with less, which is a real opportunity for small agile businesses – and a real advantage. Larger organisations are going to have trouble adopting these technologies to accelerate because they are stuck in their ways. Take advantage of what’s available and start creating.