Schematic & layout
Models that produce circuit architecture, component placement, and routing from a written specification. Trained on real designs and held to real design rules — every net checked, every footprint verified.
Software has compilers, package managers, version control, and now models that write code. Hardware has almost none of that. Schematics are drawn part by part, boards are routed trace by trace, enclosures are modeled surface by surface — skilled work, done slowly, in tools that have barely changed in twenty years.
Avenstride was founded to change how that work gets done. We train models that do hardware engineering themselves — not autocomplete inside a CAD tool, but systems that take a design from a written specification to files a factory can build. We build products around those models. The first, Conductor, is in early access now.
AI systems that take hardware from specification to manufacturable design. Four capabilities, trained together and tested end to end.
Models that produce circuit architecture, component placement, and routing from a written specification. Trained on real designs and held to real design rules — every net checked, every footprint verified.
Enclosures and mechanical parts generated together with the electronics they house. Wall thickness, fastening, tolerance, and assembly order are part of the design problem, not an afterthought.
Every generated design is checked against its requirements on real hardware — repeatable fixtures, measured procedures, recorded results. If it hasn't been measured, it doesn't ship.
Designs resolve to complete fabrication files — layout data, BOMs, drawings, assembly instructions — that a factory can build from without a phone call.
Prototypes and experiments from our own research. What's on the bench changes week to week.
A deliberately difficult board — mixed-signal, tight clearances, an unforgiving connector — used to evaluate schematic-to-layout runs. Each revision is scored on rule violations, route completion, and how much an engineer has to correct afterward.
Five prints of the same enclosure, each generated from the previous round's fit notes. Wall thickness, draft angles, and boss placement are model decisions, checked against the drawing with calipers.
The first fully model-designed board to reach the bench. Power sequencing and clocks came up on the first pass; two signal-integrity faults did not, and both are now training data.
Avenstride is an AI research lab. We train models that do real engineering work, and we build products around them. Everything we make is tested on hardware.