Inspiration
AI has transformed software engineering through Codex and other coding agents. Hardware engineering, however, still relies on workflows that have changed little for decades. Building an engineering agent is different because every result must be explainable, grounded in physics, and verifiable.
I chose direct-to-chip cold plates because they are becoming critical as AI hardware pushes against thermal limits. They are also a great engineering problem. The physics is well understood, the geometry is parametric, and the design space is far too large to explore by hand.
That is where Hexkey comes in. It generates, simulates, and validates direct-to-chip cold plate designs end to end from a natural language prompt.
What it does
Hexkey turns a prompt such as “Cool a 1200 W accelerator with water at 30 °C” into a validated cold plate design.
The copilot gathers only the information it needs, derives the full design, and presents it as an editable engineering report. Once approved, it generates the CAD model, runs an OpenFOAM simulation, validates the results, and produces a report with the design and simulation outputs.
The underlying physics is steady conjugate heat transfer. In the solid,
$$ \nabla \cdot (k,\nabla T) + \dot{q} = 0, $$
and each design is evaluated using
$$ R_{th} = \frac{T_{\max} - T_\text{in}}{Q}, $$
while balancing the pressure drop across the channels.
How we built it
I built Hexkey as a modular monolith with a FastAPI backend, a sandboxed OpenFOAM worker, and a Next.js frontend. CadQuery generates the geometry, OpenFOAM solves the physics, and every result is validated before it is shown.
GPT-5.6 is only used where language helps. It turns user intent into engineering requirements, explains the design, and suggests fixes when a design is infeasible. The engineering itself is deterministic. I used Codex throughout development to build and connect the backend, solver, geometry engine, frontend, and tests.
Challenges we ran into
The hardest problem was making every generated geometry valid enough for meshing and simulation. Small geometric changes could break the entire workflow. I also had to make sure the AI never claimed a simulation had run when it had not. Finally, CFD is slow and can fail, so the application had to handle long running jobs, cancellations, and multiple revisions without losing work.
Accomplishments that we're proud of
Hexkey turns a plain language prompt into a real engineering workflow instead of a generated answer. Every reported value comes from CAD generation, simulation, and validation. The system is fully containerized, reproducible, and designed so every run can be traced back to real evidence.
What we learned
The biggest lesson was that AI should not replace engineering. Its role is to make deterministic engineering software easier to use. Keeping the physics inside trusted tools and using language models only for communication made the system both more capable and more reliable. Building the project also showed how much faster Codex makes it to develop and maintain a large engineering codebase.
What's next for Hexkey
The next step is supporting more cold plate geometries, adding optimization across the design space, and expanding to more advanced thermal models. Beyond cooling, the same architecture can be applied to other engineering problems.
Log in or sign up for Devpost to join the conversation.