The microchips powering everything from smartphones to life-saving medical devices face global shortages, largely because designing them requires months of painstaking manual work by highly specialized engineers.
Pradeep Pydah is convinced that AI is a crucial tool to break up that bottleneck. The CEO and founder of AsterQuanta has built a platform to speed up critical tasks like circuit design and sizing, allowing analog designers to focus on analysis of the design, with the help of Amii’s expertise in reinforcement learning (RL) which learns to recognize patterns through experience.
“The future of AI is not large language models like in ChatGPT. The future of AI is actually reinforcement learning-based complex engineering problem solution discovery,” says Pydah.
“The future of AI is not large language models like in ChatGPT. The future of AI is actually reinforcement learning-based complex engineering problem solution discovery.”
Pradeep Pydah
Co-founder, AsterQuanta

A Sizeable Challenge
“We need extremely skilled analog designers with lots of experience to design circuits. It's manual, iterative in nature and complex,” Pydah says.
“There are too few semiconductor engineers, too many parts and human design capacity cannot grow easily. That is the problem.”
Circuit sizing is a vital part of designing an analog circuit. Once a circuit is designed, specialized engineers then have to work on the sizing: figuring out the dimensions of the circuit and its transistors, setting voltage baselines, and other important configurations. The decisions these engineers make have a significant impact on the performance of the device.
Circuit design and sizing is a time-consuming, tedious task, with a lot of trial-and-error. It’s made more complicated by the fact that engineers aren’t just optimizing for one thing: depending on what the chip will be used for, there will be different targets for things like power, voltage, heat and other variables. And changes to components could have different effects on those objectives.
All that combines to create a bottleneck. With more and more demand for circuits and limited engineers who are able to do the work.
The Power of "AI + X”: Amii Researchers Spark the Breakthrough Idea
Stuck inside during COVID in 2020 in Bangalore, Pydah signed up for an online course on reinforcement learning offered by Amii Fellows and Canada CIFAR AI Chairs Adam White and Martha White.
“I learned reinforcement learning, my background was in analog circuits. I joined them and then I saw the initial results and I was blown away. Believe me, that night I did not sleep,” he said.
A powerful demonstration of Amii’s “AI + X” mindset - pairing deep domain expertise with machine learning fluency to tackle complex, real-world problems.
Bridging his years of experience as an Electronics System Designer with experience in analog circuits, Pydah became convinced the approach could build a platform to let the semiconductor industry reduce the burden of scaling up analog circuit sizing, allowing the small number of specialized engineers to do more in less time.
“In 2023, I moved to Canada and the first thing I did was I sent out an email to Amii and the University of Alberta.”
Pairing up with Amii on a Custom AI Development Engagement
After hitting a roadblock with an older form of reinforcement learning, AsterQuanta approached Amii. Through support from Prairies Economic Development Canada, the two worked on a custom AI Agent as part of a Development Engagement.
Amii’s Custom AI Development Engagements is a 12-14 month engagement that leverages and ultimately transfers our expertise, technology and talent to a client. Amii hires a dedicated Machine Learning Scientist to advance and build AI alongside a client team. They also gain access to Amii’s broader research ecosystem and infrastructure while retaining 100% of the IP of whatever is built.
AsterQuanta initially started using an older form of reinforcement learning called Deep Deterministic Policy Gradient (DDPG). DDPG is designed for a "continuous action space"—meaning it handles smooth, infinite ranges of numeric values, similar to using a dimmer function for lighting instead of an on/off switch, you can “slide” to any precise point within a data set.
While this method allowed the AI to fine-tune precise decimal numbers, like adjusting exact voltages and transistor widths along a continuum, it had a major limitation: it treated the circuit as a flat list of targets while completely ignoring the circuit's structure. Because the model couldn't "see" how electronic components were wired together, tweaking one variable often triggered unpredictable, cascading effects, making the process unreliable.
Mapping the Circuit: How Amii’s Advanced RL Approach Accelerated Success
So Amii and AsterQuanta set out to create an agent using a newer reinforcement learning approach that could understand the structure of a circuit, says Payam Mousavi, the Amii Lead Machine Learning Scientist who worked on the project. To do that, they first had to translate that structure into a form the agent could understand.
“Whenever you do any kind of RL project, you need to represent your problem in a data structure. For circuits, obviously maybe you would just guess that, you know, circuits look like graphs. So why not use [them]?” he says.
The teams converted the circuit schematics into interconnected node-and-edge maps: components became nodes while wires became edges. They then used a specialized graph convolutional network that could read the graph, along with voltages, the performance targets and the measurements of the components. After several iterations, the team found the right mix of information about the structure and voltages needed to increase the agent’s performance.
Mousavi built their new agent with Soft-Actor Critic (SAC) RL, a state-of-the-art method designed to encourage exploration. Instead of just rewarding the agent based on the eventual solution it comes up with, it is also rewarded for occasional unpredictability and trying new things. This way, the agent doesn’t just get stuck in a rut following the first answer it finds; instead, the agent continues exploring new options, perhaps finding even better answers.
From Weeks to Hours: Impressive Results
The final results were impressive. Amii tested the new RL agent on a 500 mA regulator - commonly found in smartphones and medical equipment - as an ideal, real-world benchmark. They compared its performance to NSGA-II, a multi-objective evolutionary algorithm that is commonly used in problems like circuit sizing.
In as little as five hours, the RL agent was able to simulate circuit-size solutions for multiple targets, compared to 24 hours for NSGA-II, in addition to having a lower error rate.
The impressive results showed that the team was right about needing to turn the circuit schematics into graphs. The agent performed much better with increased knowledge and, since the design process is iterative, engineers adjust and tweak the variables over and over to optimize the circuit sizing — that time savings quickly adds up.
“Typically it would take an experienced designer two or three days. If it's a junior designer, two or three weeks. But an agent would take less than 10 hours. It’s not about replacing the engineers, it is about making them more productive” Pydah says.
"You cannot have AI solving the human design capacity problem, but you can have AI augmenting the human design capacity."
Scaling Success: Building Replicable Agents with Transfer Learning
The RL agent isn’t just speeding up the sizing of individual circuits, it’s creating a reusable framework that carries lessons forward to solve entirely new circuit challenges faster and more efficiently by integrating transfer learning. The team built a dual-level graph network into the system: the smaller level analyzes the finished solution, and then passes it up to the higher level, which compresses it into a sort of summary. The agent can then use that summary as a lesson that will help it to create solutions to future circuit-sizing challenges. Taking it even further, AsterQuanta isn’t creating the circuits themselves. Instead, building off the work done with the Amii team, Pradeep has been able to build a design platform that AsterQuanta’s clients can use to build their own automated sizing RL agents.
So unlike competitors that only sell the AI-design agent itself, AsterQuanta gives its clients the ability to build their own agents, offering them greater control. Pydah says they are already working with universities and global tech companies in several countries who are interested in the platform.
Empowering Global Self-Sufficiency: The Road Ahead
And Pydah says the technology is coming at a crucial time. The global semiconductor industry is highly concentrated in just a few global regions. With increased concerns over information protection and data sovereignty, many places are looking to become more self-sufficient when it comes to circuit design - democratization and decentralization of AI. And AsterQuanta’s Agent is perfectly placed to meet this demand by empowering organizations to build and control their own custom agents rather than relying on external solutions.
A Lasting Partnership with Growing Impact
While the company is now focused on putting the RL-based system to work, Pydah says AsterQuanta continues to keep a strong relationship with Amii. Embodying the core goal of knowledge and talent transfer within the Prairies Economic Development Canada engagement, the company has hired the Amii Machine Learning Resident that has worked on the project as a full-time employee.
Further, they’ve sought and secured a world-leading researcher in Amii Fellow and CIFAR Canada Chair Matt Taylor as an advisor to their Alberta team.
As AsterQuanta expands its footprint globally while deepening its local roots, the stage is set for even more groundbreaking advancement at the intersection of AI and semiconductor design.
Work with Amii
Talk to our team to find out how Amii can help your business with advanced AI adoption.
