Spice_Wizard

Analog circuit design is a bottleneck: unlike digital, it still depends on scarce senior engineers reading datasheets. LLMs can help, but they produce plausible text, not verified engineering a circuit can be mathematically correct and physically impossible. In one real run, GPT-5 made the textbook-correct resistor change and the simulation still failed, because the part's slew rate can't deliver that gain at that frequency. Only a simulator catches this, so we made the simulator the gatekeeper. Spice Wizard retrieves a known-good netlist from 4,032 indexed ADI application circuits, adapts component values using Qwen3 running on an AMD Instinct MI300X (PyTorch/ROCm), then verifies the result in real LTspice measuring gain, −3 dB bandwidth, phase, and settling against the spec with pass/fail margins. Failures feed measured evidence back for a retry. A best-of-N Candidate Arena validates topology invariants, ranks candidates by measured results, and exports an evidence bundle. Every verified pass is logged, building a simulation-labeled dataset for fine-tuning on the same MI300X a self-improving flywheel.