Chromoly

Spent the last few weeks beta testing around 20 products built by other founders.

What surprised me is that the biggest problems were rarely the product itself. One app wasn't available in my region because the developer had accidentally limited App Store availability. Another collected health data but both the Privacy Policy and Terms links were broken.

One claimed "no account required", but you couldn't actually use anything without signing in. Another had a demo where you could switch between the parent and child view, and both showed exactly the same screen. The whole product is built around parents approving what their kids do. And one failed to upload a completely normal iPhone photo with nothing more than "Failed to upload file", while other images worked fine.

None of that was a coding mistake. A region setting, a placeholder link, a sentence in a launch post, a demo nobody had updated, an error message that swallowed its own reason. The products were fine. The things around the products were not, and that is what people hit first.

Worth an hour of your week: open your own product the way a stranger would, from the link you posted on a device that has never seen it.

I'm building Chromoly, a new type of AI builder. We're still a long way from beta, but am looking for early users to help shape the product. No pressure to try but I'm happy to try yours either way.

A commenter said the core of what I'm building, working out what a change will break before it happens is impossible. He never explained, and never came back.

It bugged me enough that I went looking for what he might have meant. The strongest version is Rice's theorem: you can't write a program that reliably works out the behaviour of arbitrary programs. Applied to me, you can't look at software and compute what a change breaks. Not hard. Impossible. Settled since 1953.

If that's what he meant he was right and it still doesn't apply.

The theorem is about arbitrary code that can do anything. My system never touches that. The AI writes to a closed structured spec and the app is compiled from it so finding what a change touches is walking a graph, not analysing a program. Same reason your spreadsheet knows which cells to recalculate.

The cost is real. We can't build anything. We focus on internal tools and AI automations.

Lesson: make the strongest case against yourself, especially when the person attacking you won't.

Has anyone here hit a hard technical objection to what you're building, and did it change the product or just the way you explain it?

Building this full-time with a small team, early access opens January. Since AI app builder now describes forty products, here's precisely what this one is including what it doesn't do.

The problem it exists for. Every AI builder generates an impressive version one. Then your business changes, a field renamed, an approval threshold added, a team restructured and each change is a new AI guess against code nobody holds a map of. The most common complaints in every builder community: "it changed things I didn't mention," "it forgot last week's decisions," "I'm paying credits to fix what the AI broke."

What Chromoly does differently - the blueprint. You describe what you need in plain language. The AI doesn't write freeform code, it writes a structured blueprint: every entity, field, workflow, permission, and dependency, mapped and machine-readable. The app is compiled from that blueprint. This buys three things nothing generation-based can offer:

• Changes show their blast radius. Because every dependency is explicit, a change shows everything it touches - safe / affected / breaking before it applies. Reject the preview and nothing happened.

• The structure can't be forgotten. In chat-based builders, your app's architecture exists only in the chat history and the chat history runs out. A blueprint doesn't.

• Nothing fails silently. Every workflow step has run history and failure alerting. A broken automation tells you not a customer, nine days later.

Maintained for life. The platform owns the runtime: hosting, monitoring, security patches, AI model migrations, integration updates. When Slack changes its API, that's our job. This is only economically possible because of the compiler. We maintain one runtime all systems share, not thousands of bespoke codebases.

AI inside your systems. Workflow steps can read, sort, judge, and draft. Score a lead, extract invoice fields, draft a reply. Humans and rules decide, low-confidence output goes to a review queue. AI never sends anything and never makes decisions about people.

What it deliberately doesn't do. Marketing sites, mobile apps, consumer products at large scale, arbitrary custom code. Internal tools and AI automations only - you can't promise maintained-for-life at build-anything width. If you want code to own, Lovable and Cursor are better choices.

Pricing. A permanent free plan (build and preview, no card), paid plans from $59/mo to deploy. One credit currency - generating, editing, AI steps, runs, no expiring dual-credit systems.

Happy to go deeper on any piece, the compiler design and the "what happens when the AI is wrong" flow are the parts people usually ask about.