Inspiration

A customer can describe a custom electronic product in a few sentences, but translating that request into a manufacturable order and a robot workflow takes many separate decisions. Requirements, available components, assembly constraints, robot programming, and verification often live in different tools.

We built Uni to bring those decisions into one shared workspace. Our focus is high-mix, low-volume electronics assembly, where every variation can create new engineering and coordination work.

What it does

Uni connects customers, engineers, and cell operators around the same versioned order.

Through a conversational interface, users can clarify requirements, review structured proposals, and develop an assembly plan. The workspace brings together product information, model and tool activity, robot telemetry, camera views, gripper-design revisions, and simulation.

The central idea is traceability: each proposal, confirmation, and review should relate to the exact order and configuration being discussed.

Our prototype demonstrates the software workflow and simulation components. Completing a physical assembly and independently verifying the finished product remains the next validation milestone.

How we built it

We built a shared platform with a Rust backend, a Next.js and React web application, a SwiftUI macOS application, and a GTK4 Linux application. Typed interfaces and shared schemas connect the clients, AI services, and robotics components.

The AI layer uses structured prompts, tool interfaces, and application-level validation to turn conversation into explicit proposals and plans. Prompts include the authorized workspace context, available information, and unresolved constraints. Changes are checked against the relevant order revision.

Our documented robotics model stack includes NVIDIA Nemotron 3.5 Lightning 30B-A3B with W4A16 quantization, while Nemotron 3 Diarization supports speaker-activity processing.

We also worked with Nebius Token Factory, Nebius Cloud, and Tavily as part of our AI development workflow.

For the robotics experience, we connected live observations and camera views with a simulation workflow using NVIDIA Isaac Sim. Human review and local execution controls remain essential parts of the design.

Challenges we ran into

The hardest challenge was keeping the conversation, order state, and robotics context consistent across different applications.

A fluent model response is only useful when it respects the available parts, missing information, and current order revision. We had to make uncertainty visible and prevent an AI proposal from being treated as an approved physical action.

Integration also exposed practical problems: stale application versions, disconnected services, camera access, reconnect behavior, and differences between local and hosted environments. Testing the complete user journey helped us identify issues that isolated component checks missed.

Accomplishments that we're proud of

We connected conversation, versioned orders, structured planning, robot observations, gripper-design review, and simulation in a shared workflow.

We made the experience available across web, macOS, and Linux, and verified important paths through the actual applications.

We are especially proud of making the system inspectable. Users can review proposals, tool activity, revisions, and supporting evidence instead of receiving an unexplained instruction.

What we learned

Physical AI requires more than a capable model. It requires dependable interfaces, explicit state, clear permissions, and a way to distinguish an assumption from an observation.

We learned that model selection must account for memory, deployment constraints, and the task being performed. Prompt engineering becomes more useful when it is paired with typed tools and validation.

We also learned that a successful simulation and a working software workflow are separate milestones from a verified physical result.

What's next for Uni

Our next milestone is a complete, supervised order-to-assembly demonstration with independent product-test evidence.

We want to strengthen perception, qualify reusable robot skills, and generate order-specific robot programs through controlled tool interfaces.

We also plan to evaluate model quality, latency, and resource use on a repeatable set of representative tasks, then use those results to improve the balance between cloud reasoning and local execution.

Our long-term goal is to make custom manufacturing easier to configure, easier to review, and more reproducible.

Built With

  • nebius
  • nemotron
  • nvidia
  • ros2
  • tavily
Share this project:

Updates