In short:

What are we building? A frontier robotics system for repetitive manual work, starting with electronics assembly.

Execution Plan: Step 1: Build a SOTA (best) robotic model that allows users to deploy robots themselves - Done. ****Step 2: Grow the revenue flywheel by selling inference and robots. Step 3: Grow the data flywheel by collecting data from deployments.

Business model: Selling inference: $1/hour or a $5k license per robot

Progress: We made $43k in sales in the first week of selling at full price, since July 4, 2026 We have ~150 customers on our paid waitlist.

Demo:

Team

Igor and Viktor ran an electronics manufacturing business for 9 years, while making software and hardware side projects in parallel. Initially, we assembled products by hand ourselves, then hired and trained workers to do it.

With this experience, a team of 3 was able to build a state-of-the-art model (checked on the benchmark), production-ready hardware (we already sell it at full price), highly optimized inference (requiring a 5x cheaper local GPU), and present the product on social media (10k likes, 1M views total on launch video).

Screenshot 2026-04-20 at 13.26.56.png

Igor (CEO)

Before: Built an app that reached 1.3 million daily users.

Screenshot 2026-04-20 at 13.34.20.png

Viktor (CTO)

Before: Founding engineer at a telepresence robot company.

Screenshot 2026-04-20 at 13.27.20.png

Denys (Founding engineer)

Before: A founding engineer at K-Scale (YC W24 humanoid startup)

Assembly out workers we did

Telepresence robot built by Viktor

How it works

Large robotic models, whether VLAs or World Models, are poorly suited for repetitive tasks because they are not reliable.
We solve this by giving users the tools to train an “application model” specifically for their task. Users collect data on site, using their own materials.
We don’t use foundational models at this stage, so the model doesn't hallucinate and stays predictable and reliable. With a small dataset, the model can be retrained frequently, enabling a continuous learning loop, so users can fix errors and cover all the edge cases by updating the model after deployment.
Datasets collected by users and evaluated in real production will become reliable building blocks for a proper foundation model that accelerates training new models.