---
title: I rebuilt my 2019 CV site so a recruiter’s AI can read it
date: 2026-11-19
summary: My CV site was a Gatsby starter from December 2019, and its branch never got another commit. The new one serves the CV as HTML, Markdown and JSON.
tags:
  - nextjs
  - seo
  - llms-txt
  - json-ld
  - linkedin
language: en
author: Rafael Carvalho Monteiro
canonical: https://rafaelcmm.dev/en/blog/cv-site-ai-recruiters-can-read/
url: https://rafaelcmm.dev/en/blog/cv-site-ai-recruiters-can-read/index.md
alternates:
  - language: pt-BR
    url: https://rafaelcmm.dev/pt-br/blog/cv-site-ai-recruiters-can-read/index.md
---

# I rebuilt my 2019 CV site so a recruiter’s AI can read it

Markdown copy of the post at https://rafaelcmm.dev/en/blog/cv-site-ai-recruiters-can-read/.

My CV site stopped changing on 2 December 2019, the day I set it up. Almost seven years later I replaced it with one that publishes the same CV as HTML, Markdown and JSON, so a recruiter's AI agent can read it without scraping the page.

## Nine commits in one day, then nothing

The first commit, on 2 December 2019, imported a Gatsby CV starter. The branch has nine commits, all from that same day, and then nothing. It ran Gatsby 2.4.2 with React 16.8.6 and styled-components, and the whole CV lived in one JavaScript config file.

Deploying was a shell script. The source sat on a `development` branch, and `scripts/deploy.sh` did two things:

```bash
yarn build
git subtree push --prefix public origin master
```

The built site went into git and up to `master`, a branch that held nothing but build output.

## What runs now

The new site is a Next.js app on Vercel, at rafaelcmm.dev. It speaks English and Brazilian Portuguese, under `/en/` and `/pt-br/`, and the root redirects to `/en/`. Most pages are prerendered.

There are no analytics, no cookies and no runtime calls to third parties. Even the fonts are downloaded when the site is built, not when you open it.

## I didn't retype my CV

An import script reads my LinkedIn data export instead, and it imports only seven files: `Profile.csv`, `Positions.csv`, `Education.csv`, `Skills.csv`, `Languages.csv`, `Certifications.csv` and `Courses.csv`. The same archive holds messages, connections, e-mail addresses and phone numbers, and the script never reads them. It reads the archive in memory, extracts nothing into the repository, and the zip itself never gets committed.

Inside those files it reads only named columns. From the profile it drops the address, birth date, zip code and a few other fields before anything is written.

## Where the export is wrong

My LinkedIn profile still says Fortaleza. I live in Curitiba.

So an `overrides.json` file holds corrections like that, plus the Portuguese text, and an override wins everywhere the CV appears. Run the import again with no profile change and git shows an empty diff.

## One CV, four files a machine can read

- `/en/resume.md` and `/pt-br/resume.md`: the whole CV as one Markdown file with YAML front matter. You can [download the English one](https://rafaelcmm.dev/en/resume.md) from the nav.
- `/resume.json`: the same data in the [JSON Resume](https://jsonresume.org/schema) schema, for parsers that expect it.
- `/llms.txt`: a paragraph on who I am, then links to the CV, the posts, the feeds and the sitemap.

About `llms.txt`: as far as I know, none of the big answer engines has said it reads it. I publish it anyway, because it's cheap and an agent pointed straight at the site can use it. I just don't count it as SEO.

The CV and every post also have a Markdown version built from the same data, so the HTML and the Markdown can't disagree. Each post's HTML points to its own Markdown copy with a `Link` header, and every other page points to the Markdown CV. The Markdown is served as `text/markdown; charset=utf-8`.

## One function writes all the JSON-LD

The pages carry JSON-LD: a `Person` inside a `ProfilePage` on the home page, `BlogPosting` on posts, and `SoftwareSourceCode` for ai-dotfiles. All of it goes through one serializer, `safeJsonLd`. A `</script>` hiding in some imported text would close the script tag early, so the serializer escapes `<`. Nothing else in the code writes `JSON.stringify` output into the HTML.

Two more rules: nothing is marked up that the page doesn't show, and no field is emitted empty.

## A crawler that never runs JavaScript still gets everything

The home page has two tabs, Trajectory and Open source. Both panels are in the server-rendered HTML, and the inactive one is hidden with the `hidden` attribute instead of being removed. A crawler that never runs a script gets every fact.

The timeline has a parallax effect on the logos and years. It's a CSS scroll-driven animation, so it costs no main-thread JavaScript, and it stops under `prefers-reduced-motion`.

`robots.txt` allows AI crawlers, for training and for retrieval, and says why in the file itself.

## What I haven't measured yet

I haven't run Lighthouse against the live site, so there are no scores in this post. If you'd rather check the claims above yourself, the [Markdown CV](https://rafaelcmm.dev/en/resume.md) and [`/llms.txt`](https://rafaelcmm.dev/llms.txt) are one request away.

This site started from a written plan that I approved before any code existed; how that works is in [Spec-first with AI agents](https://rafaelcmm.dev/en/blog/spec-first-with-ai-agents/). The agent setup behind it has its own post: [How I keep one AI setup for Claude Code, Copilot and Cursor](https://rafaelcmm.dev/en/blog/one-ai-setup-claude-copilot-cursor/).
