HILOS OS: A Technical Review

HILOS OS is an industry-specific OS for ManufacturingAI that indicates the kinds of product capabilities these systems unlock, in this case specifically with footwear.

Background

Generative AI models are progressing from generating 2D images to 3D models, enabling them to begin learning not just how to represent design direction in 3D space but how to incorporate manufacturability into their results. The short-term impact of this progression is faster and more accurate product development as design for manufacturing (DfM) logic is automated into the design process. In the long-term, we believe this progression will see a shift from vertical to horizontal and composable manufacturing models - just like we see today with software development - that emphasize learning and integration over control and standardization.

Manufacturing applications each have their own specific constraints, criteria, and logic: how shoes are designed to be made varies greatly from apparel or eyewear. The logic for each application is learned and evolved over a career and varies significantly across individuals and companies. It’s also always changing with the introduction of new materials and manufacturing technologies. Teams perennially struggle to define and update this logic or to leverage it during the design process.

ManufacturingAI treats CAD as code, training AI models on engineering data in order to synthesize a set of manufacturing constraints that are continuously improved and evolved. 3D assets can be generated, updated, and improved in a networked, modular approach just as quickly as code. Product generations that normally take 2-3 years to develop, test, commercialize, learn, and update a given product line are set to shrink to weeks and months while delivering new design and performance capabilities to creative teams.

In the first weeks and month of the COVID pandemic we caught a first glimpse of this. CAD files for PPE were open sourced across the web, downloaded, printed, improved, and shared by a global community. One improvement in New Jersey was immediately felt in California, delivering product improvements at a blistering pace.

That same approach is now being seen in AI for robotics, where what one machine learns from a cycle of movements can be deployed across the fleet in real-time, accelerating group learning and rapidly improving robotic automation.

This approach is most extreme in Ukraine, where drone manufacturing mirrors this horizontal, interoperable approach and where product generations are being launched every three weeks.

Similar to the shift from vertical to horizontal that Microsoft brought to computing in the 1980s, digital manufacturing is entering a Cambrian moment that will deliver an explosion of new technologies and applications revolving around operating systems that are generative and composable while being hardware and application agnostic.

Tech Stack

HILOS is building an industry-specific OS for ManufacturingAI, adopting a horizontal approach that focuses on the specific logic for translating design inputs for footwear into actually manufacturable shoes that can integrate across a range of technologies.

Our OS is frontend and application agnostic, allowing for a wider diversity of hardware integrations and software applications, where improvements in one layer of the stack contribute to new capabilities and applications elsewhere.

1. Application layer

Standalone applications built on HILOS OS can be cloud-based or local, on-prem or off, and incorporate custom tools and third party plug-ins and integrations. Interplay is our own application and defined interface for those who may desire custom tools, models, and workflows but do not need a custom frontend, built specifically to incorporate the flexibility of custom tools and third party integrations.

Developer platform

Our SDKs allow third-party developers to build custom plugins at the application layer, rapidly expanding capabilities and features across the platform. This development community can create proprietary tools only available to a specific customer or openly available tools and plug-ins that are available for free or low-cost to the entire community of users.

2. Operating system

Our composable OS acts as a data unification layer that gives teams the freedom to create any footwear product with new or existing data in a single location. This source of truth provides teams the power to orchestrate and manage how product is designed and manufactured within the frontend experience.

3. Virtualization

Our serverless architecture allows developers to build and run applications without managing server infrastructure. In this model, cloud providers handle server provisioning, maintenance, and scaling. Applications are event-driven, executing code in response to specific triggers, and operate on a pay-as-you-go basis, where costs are incurred only when the code is running. This approach simplifies development and resources are automatically scaled to meet demand.

4. Model layer

Our Model layer combines general LLMs with specific language models trained on footwear design and manufacturing data, allowing for the rapid creation of custom models based on proprietary data streams. Proprietary data and hyperlocal models can be kept on-prem, ensuring ownership of all core IP and resulting AI capabilities.

Deep Dive into Operating System: MagiCAD

• MagiCAD

  Our underlying language for 3D asset creation, built over nerbs, mesh, and solid geometry kernel libraries. Language is an apt metaphor for how MagiCAD works: it classifies CAD assets as primitives, each made up of arguments (such as last, biteline, outsole curve) and properties, such as the color or material of that manufacturable asset. Primitives act as words that behave according to their own underlying grammar and syntax, allowing them to understand which other words they go with, how, and what that in turn creates.

  

  From a machine learning perspective, this linguistic foundation becomes even more powerful, as each parameterized “word” not only tells us how components fit together, but also translates those components into quantifiable data for analysis. By encoding 3D geometries as bits, numbers, and vectors, MagiCAD enables the extraction of feature sets ripe for training predictive models, automating workflows, and supporting intelligent design recommendations. This data-driven approach bridges the gap between geometry and computation, allowing for pattern recognition across entire product lines, optimization of material usage or performance metrics, and even the generation of novel designs through generative adversarial networks or other AI-driven techniques. In turn, this synergy between structured parametric data and machine learning accelerates innovation while preserving the intuitive, linguistic logic of the original 3D asset creation process.

• APIs

  Model agents then work to translate these words and sentences into discrete manufacturing technologies via APIs. This translation would be prohibitive without the use of agents that can take structured and unstructured data and learn from the translation process across different formats and manufacturing techniques.

Deep Dive into Application layer: Interplay

Interplay showcases our approach to generative, modular, and composable applications built for ManufacturingAI. The application is built as an infinite canvas waiting to be filled with different nodes, each one representing a different function or tool developed over HILOS OS. Designers can start with existing workflows or build their own, chaining various nodes together to deliver powerful, new product capabilities.

Nodes follow a general classification:

1. Hooks: These bring in outside data, either manually letting a user bring in an image or sketch, or automatically using webhooks or APIs to bring in real-time customer, testing, or cultural data. This includes custom AI models, or avatars, trained on both structured and unstructured data.

2. Generators: These generate assets, be they visual renders, cut views, or high-res 3D files for manufacturing.

3. Modifiers: These introduce specific capabilities to the user, modifying assets accordingly. Examples can be red-lining an existing render to change it, generating custom textures that load into a 3D texture pen for modifying a 3D asset, or changing lasts and morphing the assets accordingly.

With a modular approach, these nodes can be assembled into any order the user desires and are infinitely compostable. This has been uniquely well-suited to both the flexibility and customization designers demand while also delivering a developer platform to easily build and launch new nodes and capabilities.

Custom Model Building

This feature within Interplay is surprisingly powerful, allowing users and teams to quickly build custom models with a small amount of sample data or with webhooks and APIs continuously bringing in new data.

By introducing multiple model types—such as generative models for style exploration and surrogate models for performance optimization—Interplay enables a truly adaptive, data-driven design process. Image libraries, for instance, can be leveraged to train generative networks that produce new footwear concepts aligned with a brand’s established aesthetic or functional requirements. Meanwhile, test fit measurements and performance data can feed into surrogate models that predict how various design tweaks might affect comfort, durability, or other key metrics, streamlining the prototyping process. By combining these specialized approaches under one cohesive platform, Interplay empowers teams to continually refine their footwear designs with high fidelity, harnessing the user’s existing data to unlock faster innovation and usher in more personalized product experiences.

To learn what manufacturing AI means for the future of footwear, read on below:

Or if you’re an executive at a large brand exploring how you can bring this into your company, we have a few tips for you below:

Comment and subscribe below, or reach out directly.