GenieEDA

• The idea: free, local-first tools for designing circuit boards, and engineers for hire when a team needs the board actually done. Software as the funnel, a service as the business
• What existed: a browser workbench that really edits schematics and boards, with reviewed, undoable changes; a marketing site, pricing, sales PDFs and a launch film
• What happened: no clients. Looking back at the repository, it is easy to see why: nobody could have become one

Problem

Small hardware teams (embedded, robotics, IoT) lose weeks to board design that isn't their product. The tools lock a design into one vendor's file format, the reasoning behind each change lives in someone's head, and "the design-rule check is green" gets mistaken for "this board is ready to fabricate." In the summer of 2026 I wanted to sell the fix as a service: scope a board, engineer it, hand back an open project file with the evidence for every decision.

Solution

GenieEDA was three tiers. genieeda, a free workspace that runs in the browser with no account, keeps one open project file, and turns every edit into a reviewed, journaled diff. genie engineering, paid board design: a pilot board for $5k–15k, team capacity at$25k–50k a year, every number marked as an initial hypothesis. And genie lab, an office fab-lab kit with a robot, which was an idea and a page.

What was pitched next to what existed: the workspace's editing, routing, DRC, requirements, BOM and reviewed change proposals worked; KiCad import and the release gate were library code only; the engineering service and the lab were pages. Intake switched off, robots.txt disallowing everything, 0 of 3 outreach sequences approved, 0 boards made, 0 clients.

The workbench part is real. Opening the example CAN-node board gives a working schematic editor with 29 objects on 8 nets, its own structural electrical-rule check (labelled honestly as not KiCad's), undo and redo, and an inspector:

The GenieEDA workbench with the illustrative CAN node schematic open: an MCU and a CAN transceiver wired on a sheet, a tool palette, a symbol search, an inspector, and the nets list

How

Almost all of it was built in about fifteen hours on 27–28 September, and mostly not by me typing. I wrote a 77-task backlog and a master prompt, and Codex and Claude subagents worked through it in parallel, about seventy-five git worktrees, 356 commits. I reviewed, merged and steered.

• Code: a TypeScript monorepo of about 117,000 lines in 25 packages, plus a Rust/WASM polygon kernel for board geometry
• Apps: the website, the browser workbench (IndexedDB storage, a service worker, works offline), an unsigned Electron shell, and a pipeline that renders the launch film from scripted, real-time captures of the workbench
• Hosting: one Cloudflare Worker serving static assets, with 17 D1 migrations for identity, sync, jobs, reviews and an engineering portal, almost all of it behind a "planned release" flag
• AI: six typed tools an agent could use to propose edits, each applied only after a person approves the diff. They were tested against a scripted fake model, never a real one

Results

Zero clients, and the repository explains it better than I can. The intake form was switched off (INTAKE_ENABLED=0). robots.txt disallowed every page and the footer said "not indexed, no analytics", so nobody could find it. All three outbound email sequences sat at approved: false. The disclosures page said, accurately, that no engagement had been delivered and no legal entity or contact address had been published. No board was ever fabricated.

Lessons

I built the product and the brochure and skipped the only part a business can't skip: asking someone to pay. Agents made it cheap to produce a convincing-looking company in a night, which made it easy to keep polishing the company instead of talking to a single hardware team. The workbench is a decent narrow tool. The next time I try to sell something, the first artifact will be a conversation, not a pricing page.