This week I launched the AI resume optimizer I built as a physician — with a rule most AI tools refuse: it cannot invent anything.
Every resume gets read twice: once by software, once by a human. Software goes first, in seconds, and it punishes formatting that buries the words it scans for. That's how "optimize my resume" became an entire product category. The problem is what most of those tools do: they fabricate. They add skills you never listed, inflate titles, invent metrics. Recruiters smell it, and the systems that screen candidates flag it. A fabricated resume is worse than none — it follows people from one application to the next.
So this week applystudio.app went live with the opposite rule baked in. The AI rewrites your bullet points to match a job description's keywords, but every claim must trace back to something you actually did. Every change is shown before and after, because nothing should happen invisibly to the person whose name is on the document. Output is ATS-friendly by design — plain, parseable, honest.
The stack is the boring pattern I keep returning to: Next.js and FastAPI on a Proxmox box already in my house, PostgreSQL, an async job queue, Ollama for the model. A second product on my own hardware, no new API bill. An AI coding agent wrote most of the implementation; I did the review, the product calls, and the refusal logic — because the refusal logic is the product.
Now the honest part. There is no launch-day spike for a resume tool, and I'm not going to pretend there was. The plan is a grind: thirteen SEO posts about the resume problem were live before the product page existed. Directories are listed where people search competitor names. A free tier tests itself. None of it is exciting and all of it compounds. Last week I launched a public API and watched a trickle — I'd rather show you a real trickle than a screenshot of a dashboard I invented.
The throughline is the same one I use at work: evidence over hype. I check claims for a living. Software that is structurally unable to lie is that discipline in another language.
Tools I leaned on: Next.js + FastAPI (the app), my Proxmox LXC (the server), Ollama Cloud (the model), Stripe (billing), and the AI coding agent that wrote most of the lines I reviewed.
Next week: the writing tool I built because I kept losing track of my own fantasy world — and why authors trust software that remembers.