The industry is currently stuck in binary debates: Local vs. Cloud, Open vs. Closed, or even AI vs. non-AI. The reality is somewhere in between. Just because a model can be run locally today doesn't make it useful for most people; raw inference isn't a functional product and weights alone is not intelligence.
We need to bridge the gap between "here is an open model you can run on expensive hardware" and "here is how to get more from your existing devices." Just as NeXTSTEP built the foundation that redefined personal computing at the software and OS levels, we need a new foundational layer for the intelligence era.
Even as model capabilities improve for the edge, we still need cloud infrastructure for certain tasks. We need edge-first hybrid systems that deliver privacy where it matters and power where it is needed. However, local-first hybrid AI today is missing the core engineering foundations required to make it accessible, trustworthy, usable, and truly personal.
Nua is an AI engineering and product lab currently focused on building these missing foundations as a set of composable primitives at the intersection of three key areas:
Hybrid Edge-Cloud Foundations: Building optimized runtimes and kernels, reliable orchestration and tool-calling on the edge, and other infrastructure that is not just functional but also efficient.
On-Device Personalization: Engineering ways to enable ongoing learning to adapt to user preferences in resource-constrained environments. Personalization shouldn't require a trade-off with privacy or data sovereignty or portability.
Interactions: Moving beyond chatbots to explore new HCI primitives like drawing or embedded interactions that are context-aware, personalized, feel native, and are accessible to both humans and AI. Some of these interactions feel long overdue, even without AI behind them.
The core architecture is inspired by the Unix philosophy, and informed by the friction of shipping real-world multimodal products and AI. While some early explorations around "Computer Use" or other multimodal interactions didn't always hit the mark, they provided further clarity and highlighted the infrastructure gaps.
Ultimately, useful products are what end-users will care about, but products are only as good as the infrastructure they sit on. The goal is to build shared primitives for key pieces of infrastructure as well as experiences; primitives and experiences that are useful, usable, and accessible to both AI and end-users.