Fort Worth, Texas · USA · Central Time · open to work
Marketing Operations & AI Automation
I build the systems behind lead generation and client websites, then automate the repetitive half. Get my résumé in your terminal, or ask your agent to fetch it.
01 Work history
Every job below was accuracy work before it was systems work — that order is the whole point.
$ git log --graph --career
(init) first commit · learning by shipping 373d1b6
02 Pillars
The same three ideas underneath every workflow below — which is why they survive being handed to someone else.
Nothing downstream is trusted more than the record it came from. Ingestion, deduplication, validation and reconciliation against a source of truth — with every value traceable back to something that can be re-checked.
Work that runs the same way twice without supervision. Bounded scheduled loops, hard requirement gates, human-in-the-loop review on anything that writes, and an SOP written before the process is handed off.
A result nobody verified is not a result. Test suites, offline smoke runs, freshness gates, real-browser verification, and a validation report shipped beside the data as the evidence.
03 Workflows
Each one runs today. Expand any of them for what it does and why it is built that way.
$ ls ~/workflows
Grade a vertical once against weighted criteria, cut it into bands, and store the grade as a lookup so later stages read it instead of re-deciding it. Every score cell carries its provenance — measured, derived, recalled or assumed — and an unverified cell is not allowed to disqualify a segment or win the budget.
Cheap HTTP for everything, headless browser only where the cheap pass demonstrably failed — and the decision is made on the response body, never the status code. Pages come from the site's own sitemap rather than a blind crawl, and any fact that will be quoted back to a person is verified against the page or visibly quarantined.
Reported figures checked back against platform data, inconsistencies surfaced before they reach a decision, and records left clean enough to hand to a client without a caveat.
16 source families scanned on a bounded daily loop with a kill switch, run through a hard requirements gate before anything is scored. The loop refuses to publish a scan that came back empty, is over six hours stale, or lost more than half its sources — a silent bad run is worse than no run.
Drafts are generated, then fact-checked against the underlying record by a second pass that returns literal find-and-replace edits. The document gets corrected rather than rewritten, so it never quietly drifts away from the source it was built from.
Repetitive data entry accelerated without surrendering accountability: the system proposes, a person approves, and nothing writes unattended. Speed is bought from the review loop, not from the accuracy.
The process written down while it is still fresh, so the next run needs no supervision and the knowledge does not live in one head. This is the step that turns a task someone did into a task anyone can run.
Cleaned data ships with a report showing exactly what changed and why. The report is the deliverable — the guarantee is proof rather than a promise, and it is what makes the work checkable by someone who was not there.
A real browser driven across three viewports, failing the build on accessibility violations, console errors, layout overflow, undersized targets, or a page that rendered nothing. Built because a green test suite had already let real defects through — including a page that scored perfectly while displaying nothing at all.
04 Operating layer
The through-line under all of it: I design the system and hold the judgment, and the machine does the volume. That is why the same person can run a Meta Ads account, a client website and a data pipeline — in each, the intensive work is automated and my job is the structure, the standards and the gate.
05 Projects
Working systems I own end to end — each one exists because a manual process was costing me hours.
$ ls -la ~/projects
External traffic partner for an AI SaaS product with no access to the client’s analytics, ad accounts or billing, so attribution was built in-house. The rig: tagged outbound links, a weekly sweep, QA and metrics loop that ranks what worked, a frame check that rejects dead footage before it reaches a posting slate, and an approval gate that creates every post as a draft and reads it back from the server before anything is scheduled. 285 commits.
A two-campaign Meta Ads system: Campaign 1 buys cold attention cheaply with video-first creative, Campaign 2 converts only the people Campaign 1 already warmed up. Over 90 days for a college client: 487,000 impressions at $0.47 CPC, 856 leads at $4.22 each, and 127 applications at $28.43 each, from an account with no prior advertising and no pixel installed. Full case study at ads.andrew-rocha.here.now.
Ongoing client work for a ministry: an AI proxy holding every API key server-side, a Dropbox-to-library autosync watcher on a scheduled task, research and hub packs, and a site remodel. A CRM data-quality audit across 2,029 contacts took customer records out of version control and added a check so they cannot return. Ships behind a definition-of-done gate that has to pass before anything is called finished. 625 commits, 1,034 tracked files.
A faith-based pregnancy journal and family keepsake app I own and ship: offline-first, no login, every entry stored on the device in IndexedDB, live on Cloudflare Pages. Includes an OWASP ASVS 5 security review with a tracked remediation plan. 460 commits.
A document pipeline that takes structured inputs — a person's details, written testimony, source references and authorised context from an existing database and document corpus — and assembles an illustrated 16-page PDF from them. A versioned data schema sits under change control, with purge-and-rollback tests for sensitive records. 290 commits and 949 passing tests, because a generator that silently drops an input is worse than no generator.
A daily job pipeline: 16 source families scanned into SQLite, run through a hard requirements gate, scored, and rendered into a self-contained dashboard. 315 commits behind three gates — a 451-test pytest suite, an offline pipeline smoke run, and a real-browser harness across three viewports that exits non-zero on accessibility violations, console errors, layout overflow, or a dead interaction.
A complete private archive of a year of coursework, built so the data survives losing access to the system it came from: 30 courses and 268 lesson pages, with grades, files, video transcripts and email pulled out of a hosted LMS, normalised, and stored somewhere retrievable. A watcher checks 13 source syllabus documents every 30 minutes and flags any change for review instead of applying it. 407 commits, 759 tracked files. This is records management as a migration problem — get it all, keep it faithful to the source, and make it findable afterwards.
A status board covering 56 projects and the client engagements at once, rebuilt from source every morning by a scheduled task: ledgers, git state and workflow registers are read fresh each run and nothing is stored in between, so the board cannot disagree with the work. 62 commits; 125 tests across 4 test files.
A private widget platform: a password-gated builder plus a public embed runtime, so a marketing site adds a form or gallery without a subscription or a third-party CDN. 51 commits and 44 test files; the embed ships as self-contained output because an embed that phones home is a dependency nobody audited.
A productized spreadsheet-rescue service: messy client data in, cleaned and deduplicated data out, plus a validation report showing exactly what changed and why. The report is the deliverable; the guarantee is proof, not a promise.
A verification gate that made a green test suite mean something. Drives a real browser at three viewports and fails the build on accessibility violations, console errors, sideways scroll, undersized tap targets, or a page that rendered nothing, the last one because an empty page used to score perfectly.
06 Skills
07 Education
08 Colophon
The same method as the work above, applied to the résumé itself: an AI-assisted build, with checks that fail the publish instead of trusting a look-over.
09 Contact
Remote from Fort Worth, Texas (US Central). Email is the fastest way to reach me: rocha@yahwaylove.com