MMIC skills and IMS 2026 demos

A skill packages one piece of RF know-how so an agent runs it the same way every time. I built skills for the MMIC components designers model over and over, and measured them against doing the same job by hand.

Skills

SkillBy handWith the skill
Spiral inductor: solve for L and Q, export a turn sweep as one MDIF file90 to 150 minabout 2 min
Interdigital capacitor: drive the fingers to a target C and report Q60 to 100 minabout 2 min

By-hand times are conservative estimates; the skill times are measured wall clock.

A square spiral inductor on GaAs in HFSS
A 2.5-turn square spiral on 100 µm GaAs: 1.32 nH, peak Q of 21.2 at 16.8 GHz. The 1.5, 2.5 and 3.5-turn variants went into one MDIF file for the circuit simulator.
An interdigital capacitor in HFSS
An interdigital capacitor, finger count swept to hit a target capacitance.

A skill chain

Chained together, the skills built and solved a three-pole GaAs MMIC bandpass filter end to end in 36 minutes for $8.12 in agent cost: MIM capacitors, spiral inductors, air bridges and vias, then the HFSS solve. The first pass gave a real bandpass response, peaking near 3.2 GHz against a 2.3 GHz design target, so the next step is tuning.

A three-pole GaAs MMIC bandpass filter built by a chain of skills
Three shunt resonators with capacitive coupling, built by the skill chain.

IMS 2026 demos

Three of the demos from the DesignForge booth at IMS 2026.

A fully parameterized Archimedean spiral, animated through its turns, width, radius and spacing.
A third-order coupled-line Chebyshev bandpass filter for 5.15 to 5.85 GHz, planned, built, solved and plotted.
A 2.4 GHz inset-fed patch antenna, sized from closed-form equations and built in HFSS.

View Next

DesignForge: A Klopfenstein Taper from One Prompt

One prompt in DesignForge AI designs, builds and solves a 50→90 Ω Klopfenstein taper in Ansys HFSS, then checks it against the analytical model. The orchestrator ran on Claude Sonnet 5.

Parallax: telling an AI agent where its drawing is wrong

A five-phase image pipeline that helps an AI agent reconstruct engineering diagrams such as PCB traces and RF filter layouts. On 13 diagrams it beat single-shot multimodal reconstruction every time, with mean geometry IoU 0.95 against 0.59.