GadgetForge

It started with a water bottle. I love my smart bottle: it tracks what I drink, filters it, nudges me to be better everyday :P It also costs more than most people will pay for a bottle. So we set out to design a cheap clip-in tracker for any bottle, and used AI to help. Every new session forgot what we had already solved and re-raised the same problems. A friend in embedded robotics said he has the same pain at work, but would never let an AI near a real build without being able to trust it and steer it. That is the problem Gadget Forge solves: an AI that designs hardware and remembers what it already proved. Describe a device. Eight specialist agents take it through physics, parts, circuit and PCB (KiCad, driven over MCP) and out comes a buildable design: BOM with cost, board files, 3D model. The model never grades itself. Every action is written to a MongoDB Atlas ledger as “sent” and only becomes “verified” when real code passes it (tests, parity checks, KiCad DRC) and a commit lands. Kill the process mid-run and it resumes at the last verified step. Change the spec and only the affected steps re-verify. Every failure is embedded in Atlas, and Vector Search hands the lesson back on the next attempt. You can watch every step live, cancel, and decide what merges. Atlas is the foundation, not a feature: queue, checkpoints, ledger, memory and live feed in one cluster. An AI designer is only as accurate as its memory, and this one stores only what code verified. Why it matters: a hardware spin takes six to ten weeks and 77% of firms can’t hire the engineers to do it. This is for the maker with one idea and no EE degree, and for the team with a hundred builds to track. Next: live supplier pricing, multi-user workspaces, materials catalogue, the remaining phases with coded gates, screening gate that presets basic build conditions.

Team