Ideas, under
physical constraints.

Build the laboratory a capable agent needs to do real engineering.

I am developing environments for agents to create, inspect, simulate, and revise RF integrated circuits. The focus is on the complete interaction with the design environment: usable tools, persistent native state, controlled execution, recoverable errors, and independent verification.

I prefer to reuse a general agent runtime and invest in the parts that are specific to engineering. The aim is not a scripted demonstration with the solution already hidden in the tools. It is a workspace in which the agent makes meaningful design decisions and leaves an editable result.

Questions I am working on

How much freedom should the tool layer expose? How do geometry and circuit decisions inform each other? What evidence is sufficient to support a capability claim? How should a human take over and continue the resulting design?

What counts as progress

A larger, clearly stated design responsibility; an independently checked result; and an honest account of cost, failure, and human intervention. The goal is useful RFIC design, not increasingly elaborate orchestration.

Use expensive physical evidence where it can change the decision.

I am interested in electromagnetic modeling, passive-device design, inverse design, and uncertainty-aware multifidelity evaluation. The objective is to make physical feedback more useful to a designer or agent—not merely to replace a trusted solver with a faster approximation.

A promising evaluation tool should expose its domain of validity, preserve the identity of the design it evaluated, and distinguish a prediction from a verified result. It should help navigate design space while leaving the final evidence requirement intact.

Questions I am working on

When is a cheap model good enough to screen a candidate? When could uncertainty reverse a decision? How can reusable simulation knowledge support a new geometry without overstating generalization?

Ask what the physical signal path can compute before treating it only as a channel.

I am exploring microwave systems in which controllable signal transformations and nonlinear responses participate in computation. Of particular interest is the possibility of using an amplifier’s nonlinear operating regime as a resource rather than treating every departure from linearity as something to suppress.

This is also where two meanings of “compression” meet. Gain compression is a physical response; information compression is a question about representation. Their coexistence is a useful starting point for an investigation, not proof that the mechanisms are equivalent.

Questions I am working on

What does the nonlinear path contribute? How robust is the trained system to changing operating conditions? Where do conversion, calibration, control, and readout costs belong in the system-level comparison?