MEDUSA

Inspiration

Defense organizations sit on piles of underutilized hardware, cameras, sensors, radios, compute, that could be repurposed into real defense capabilities, but nobody has a clear way to check what's compatible before deployment. The same gap shows up when starting a prototype from scratch: engineers need a design that respects the budget from day one, not a wish list that gets cut apart later. Boeing's SBInet border-surveillance program proved this isn't hypothetical. Public GAO reports describe cameras failing in bad weather, radar misalignment, and signal loss, expensive hardware assembled without that clarity. We wanted to know: what if choosing defense hardware, whether from what's already sitting in inventory or for a brand-new project, wasn't a spreadsheet guessing game, and the system understood your budget well enough to suggest alternatives that actually fit it?

What it does

MEDUSA turns a plain-English mission ("monitor a perimeter at night in bad weather, alert an operator") into a real hardware design pulled from actual owned inventory. It walks a graph of sensors, compute, comms, and power to propose a working system, flags compatibility issues with evidence and in-stock fixes, and shows exactly how one swap ripples through the rest of the design before a human approves it.

How we built it

100% in Jac. Inventory, requirements, and architecture live as nodes and edges in one graph. A walker filters out anything with insufficient quantity or a hard incompatibility, then uses visit statements by llm() to reason over what's left, so the LLM chooses among engineering-valid options, never around them. Every visited node and rationale is recorded for a fully explainable design trail.

Challenges we ran into

Keeping the LLM constrained was the hard part; letting it reason about mission fit while never letting it override a hard incompatibility or an inventory limit. We solved it by filtering the graph deterministically first and only then handing the walker to by llm().

Accomplishments that we're proud of

A single hardware concern, an unverified low-light camera has the ability to expand automatically into every affected mount, power, comms, cost, and test implication, and can't reach an approved design until four hard gates all pass.

What we learned

Explainability isn't a feature you bolt on after the fact. When reasoning happens as traversal over a typed graph, the "why" comes for free. We also learned how much discipline it takes to keep an LLM genuinely constrained rather than just prompted to behave.

What's next for MEDUSA

Real inventory imports from live ERP and warehouse systems, true multi-user collaboration with concurrent branching, review, and merge, and expanding past simple mission scenarios to arbitrary complex defense hardware scenarios and larger, messier inventories.

Built With

• ai
• jac