Compression Without Compromise: A Perforated Customer Spotlight

Perforated customer thoro.ai is doing some very cool stuff. Their autonomy platform enables industrial OEMs to turn their equipment into autonomous robots. From perception to navigation and safety (and a whole lot of other details along the way), we're impressed by their autonomous technology. But as you might expect from any edge application of AI, there are some familiar pain points and constraints. Optimizing for smaller model size is a perennial hurdle, as is the data burden of training those models.

Smaller Models, No Tradeoffs

Not only does the hardware limit the size of models you can realistically deploy, but larger models also run too slowly for real-time detection and avoidance of equipment, people, and other obstacles. The eternal tradeoff is that small, fast models are also less accurate. Essentially, thoro.ai is always fighting to optimize for both model accuracy and model size. By optimizing their models with dendrites, they achieved incredible results. It feels like having your autonomy platform cake and eating it, too.

Perforated ran a host of experiments, adding dendrites to various scaled-down versions of thoro.ai's model, and the leanest of those models was able to cut parameter count by 70% while maintaining (actually very slightly improving) the accuracy of the original. This unlocks major improvements to runtime and even opens up the potential to run an additional image stream on the very same hardware.

This graph shows four of the best results from dendritic optimization of thoro.ai's model. The smallest dendritic model is 70% smaller with comparable accuracy. Notably, Perforated also produced a model with a 50% size reduction and a major improvement in accuracy.

Unburdened By Data Collection

Unlike some parts of the tech industry, there's a bottleneck in collecting the data thoro.ai needs to train their models. They need site-specific data. That means an engineer flying to the site, spending a week collecting data, and then another week working with a vendor to manually label that data. Head of Software Chris Dunkers spoke on this at the Pittsburgh Technology Council's Think!AI summit this February, alongside Perforated AI President & COO Erin Yanacek, as well as thoro.ai CFO & Founder Dan Beaven. He expects that, with the addition of Perforated technology to their stack, engineers' time will be cut by 50-80%. This is because dendritic models are optimized to the point that they can achieve the accuracy and size thresholds of traditional models with less than half the training dataset.

Perforated AI President & COO Erin Yanacek speaks alongside thoro.ai's CFO & Founder Dan Beaven and Head of Software Chris Dunkers at the Pittsburgh Technology Council's ThinkAI! Summit in February 2026

The Bottom Line

Our customers are already doing pretty brilliant stuff, but at the end of the line, they're also businesses. By leveraging our novel optimization technology, companies like thoro.ai can work faster and deliver greater value to their customers at a lower cost. When you have models that are smarter and smaller, you can deliver superior results from the same hardware in ways that traditional methods just can't. We're empowering companies to turn a tradeoff into functional advantages and a competitive edge.