Know-how does not compound
Each new block repeats old timing fixes. When senior engineers leave, the lessons leave too.
Recursive Self-Improvement for Semiconductor Engineering
Infrastructure for Recursive Silicon Intelligence.
An agent writes the circuit. Chip tools grade it. Those graded runs train the next model. The loop is built to run on machines you already control, so the design does not have to leave your network.
The Problem
Where chip teams get stuck.
Generic AI does not know your flow
Public models never trained on your proprietary libraries, scripts, or past designs.
Your data has to stay private
Foundry process data comes under NDA. Design files are the company’s core asset. Neither can go to an outside AI service.
Customer Stories
Custom AI systems, built with silicon teams.
An RSI harness for accelerator design
We built a recursive self-improvement harness around the customer’s own design checks. Agents design, their tools grade every attempt, and the graded runs train a model that gets better each round. It all runs on their machines.
A model fine-tuned for their design work
We worked side by side with their engineers and fine-tuned a model on their own design data and expert corrections. The model runs inside their network and belongs to them.
Research
Science in Broad Daylight
How should an RSI system evaluate agent performance on open-ended design tasks when self-reported claims cannot be trusted?
The Goal of Science Is Not to Win
How should an RSI loop decide what candidate to explore next across vast design spaces without waiting hours for sign-off?
AI Should Build Its Own Research World Model
How can a self-improving design loop retain and query verified lessons across sessions so agents never repeat failed fixes?
Agent-Native Research Artifacts
How should complex branching design iterations and negative results be packaged so future agents can inherit and rerun them?
Beyond Natural Language: An Agent-Native Language for Autonomous Science
How can autonomous agents formalize and mathematically verify engineering hypotheses beyond ambiguous natural language?
Curie: Toward Rigorous and Automated Scientific Experimentation with AI Agents
How do we enforce experimental rigor, controls, and auditability in autonomous multi-step agent workflows?
EXP-Bench: Can AI Conduct AI Research Experiments?
Can autonomous AI agents design, execute, and debug complex multi-step technical experiments end-to-end?
Work with us.
If you’re designing silicon and want to see what recursive AI can do on your workflows, we’d love to chat.
Talk with usTeam
Frontier AI and silicon engineering, in one team.
AI & Training
Jiachen “Amber” Liu
Research scientist at Meta Superintelligence Lab. Machine learning research at MIT and Apple. Ph.D. in Computer Science from the University of Michigan.
Silicon Architecture & DV
Yiqian Min
Memory subsystem architecture for Meta’s MTIA accelerator. High-performance networking silicon at Marvell. CPU design verification on Apple A5, A6, and A7.
Our team and advisors come from
Backed by