Growth & Marketing Engineer

I build growth funnels that are measured, private and automated.

I spent 2024 to 2026 at Flex, a US fintech that lets renters split their rent into two payments. I owned the tracking, privacy consent and experimentation behind its marketing sites, and I handed repetitive work to AI agents that had to pass a check before anything shipped.

Email me

Remote from Córdoba, Argentina · UTC−3 (1–2 h ahead of US Eastern)

LinkedIn GitHub

Tools behind the work

Transcend Airgap
Google Tag Manager
Amplitude
Webflow
Unbounce
CloudFront
Athena
Snowflake
SQL
Zendesk Guide
Handlebars
esbuild
Playwright
node:test
OAuth 2.0
JavaScript
A/B testing
Claude Code
MCP
Python
GitHub Actions
Notion
pandoc
AWS Lambda
Terraform
Datadog
PagerDuty
httpx

Privacy compliance98.9%of page views loaded the consent manager, up from under 1.5%
Ad measurement21%of paid-ad visits never reached analytics, measured by a pipeline I built
Button speed0.5 sfaster to open the next page after a button click. I removed a wait built into the old tracking
AI automation36 pagesnew pages added to outage monitoring by an AI agent using my audit tool

Selected work

Six problems, what I did, and what changed

Finding the 21% of paid-ad visits that analytics never saw

Ad measurement21%of paid-ad visits never reached analytics, measured by a pipeline I built

The problem: Paid-ad arrivals and analytics page views did not line up, and the size of the gap was unclear.
What I did: I built a pipeline that recorded every paid arrival with a beacon on CloudFront, then joined it to Amplitude page views in Athena and Snowflake. The join covered 99.8% of page views, and a query cost about $0.07.
The result: Paid arrivals were 1.26× recorded page views, a gap of about 21%.
Behind it: The goal was not the pipeline. It was knowing how far the dashboard could be trusted.

Ad click
CloudFront beacon
Athena join
Snowflake
Gap measured

CloudFront
Athena
Snowflake
Amplitude
SQL

Rebranding the help center every year, and rebuilding it in 2026

Testing600+automated checks guarded the help center I rebuilt from scratch

The problem: Flex’s Zendesk help center is rebranded every year, and I did all three: 2024, 2025 and 2026. The 2026 one was not a repaint. It needed a ground-up rebuild on Zendesk’s Guide Templating API v4.
What I did: I built 32 templates with esbuild and token-driven CSS, with accessibility first, plus a chat loader that talks to the mobile app through a versioned contract. It was guarded by 328 unit tests and about 300 Playwright checks, and it went live on Oct 16, 2026 through a reviewer-gated deploy with a smoke test and automatic rollback. I also moved deploys to OAuth 2.0 ahead of Zendesk’s API-token sunset.
The result: A support site that is tested like a product and can be rolled back in one step.

Change ready
Reviewer approves
Deploy
Smoke test
Automatic rollback if it fails

Zendesk Guide
Handlebars
esbuild
Playwright
node:test
OAuth 2.0

Faster buttons, and A/B tests you can believe

Button speed0.5 sfaster to open the next page after a button click. I removed a wait built into the old tracking

The problem: The old click-tracking script held visitors on the page for a moment after they clicked a main button, so the next page opened late. A/B test counts also looked too high.
What I did: I upgraded Amplitude’s tracking library (its SDK) to a newer version that records button clicks on its own, and compared it with the old script over a week of real traffic: their click counts differed by only 0.15%, so I deleted the old script. The new engagement and error events exposed three hidden production bugs (about 3.3K unhandled errors a day) within an hour. I also fixed A/B test counts that were inflated 1.34–1.56×.
The result: Buttons responded about half a second sooner, and experiment results rested on honest counts.

Old tracking script
Amplitude upgraded
Counts differ by 0.15%
Old script deleted

Amplitude
JavaScript
A/B testing
Unbounce

AI agents with a human at the gate

AI automation36 pagesnew pages added to outage monitoring by an AI agent using my audit tool

The problem: Two jobs were repetitive and easy to get wrong: publishing legal documents to the website, and checking which pages the outage monitoring did not watch.
What I did: I created Flex’s internal Claude Code plugin marketplace (manifest checks and AI review in CI, golden-snapshot tests, opt-in installs, CODEOWNERS) and shipped two plugins on it. /legal-ticket turned a Notion ticket into a checked CDN page, a PDF and a deploy runbook. Nothing was written unless a ≥ 0.996 word-accuracy gate passed, and the production upload stayed human-only. /uptime-scope-audit compared the sitemap and the CMS API with the monitoring list and opened a pull request.
The result: The audit’s first run found 64 pages nobody was watching (the sitemap alone misses ~28%). Flex’s infrastructure agent then ran it unattended, and it added 36 pages to the monitoring list.

Notion ticket
Plugin builds page
Accuracy gate ≥ 0.996
Human uploads to production

Claude Code
MCP
Python
GitHub Actions
Notion
pandoc

Watching 479 URLs for a few dollars a month

Monitoring cost479 URLsmonitored for under $10 a month (est.), vs $2.2–4.3K on Datadog Synthetics

The problem: Watching this many URLs with Datadog Synthetics would have cost an estimated $2.2–4.3K a month.
What I did: I wrote, on my own, a serverless poller in Python 3.12: AWS Lambda containers on EventBridge Scheduler, with Datadog and PagerDuty for alerts and Terraform across five environments. It paced requests per host so Cloudflare did not block it, had 300+ tests, and a pre-launch audit of the URL list prevented about 40 false alerts on day one.
The result: 479 URLs watched for single-digit dollars a month (my estimate), against $2.2–4.3K.

EventBridge schedule
Lambda poller
Datadog
PagerDuty alert

Python
AWS Lambda
Terraform
Datadog
PagerDuty
httpx

Also shipped

More work from the same team

Flex Website 2.0 rebrand

Owned 63 pages of the Webflow rebrand (my third yearly site rebrand, after 2024 and 2025) and the 414-URL cutover inventory, then repaired the post-launch gaps (17 authors recreated, 222 CMS items re-linked, 49 legal files re-keyed).

Legal-content pipeline

Designed a gated DOCX → HTML pipeline (pandoc + BeautifulSoup, word-level fidelity gate ≥ 0.996, human-only production upload) and moved 47 of 58 legal pages to S3/CloudFront, verified on QA within two days (25/25 PDFs SHA-256-identical to the originals). Restored 231 words that had been truncated from a live bank agreement.

SEO and structured data

Fixed placeholder titles across four CMS templates and added Article schema with bylines: Ahrefs had flagged 226 short-title pages; a crawl of all 275 live pages afterwards found zero.

Billing-statement emails

Migrated two partner banks’ billing-statement emails from Iterable to Braze Canvas (Liquid), with a 36-field data contract checked against production payloads and fixes for Gmail clipping.

Visual and performance QA harness

Built a Playwright + Lighthouse harness (10 browser/device configs, Figma reference comparison, WCAG contrast and link checks). Partner pages scored 89–95 desktop and 78–94 mobile performance, CLS ≤ 0.041, accessibility 92–96; a full-site crawl (413 routes, 25,836 links) found 7 broken links.

Paid landing pages

Built the “Split your rent” calculator widget for Flex’s Unbounce landing pages, replaced jQuery carousels with vanilla JS, and caught a pre-launch regression that had disabled every overlay-style CTA.

How I work

Three habits behind all of it

Measure before you trust
A dashboard is only useful if you know how far to trust it. I measured the paid-traffic gap from raw data and got 21%.
Automation gets a gate
Agents do the repetitive work. A check decides whether the result is good enough, and a person still presses the button for anything that reaches production.
Every change can be undone
Snapshots, rollbacks and reviewer-gated deploys mean a bad release is a quick revert, not a weekend.

Open source

Code you can read

The code I wrote at Flex is private, so I rebuilt two of its ideas from scratch as small public projects.

consent-gated-tracking

A tiny browser library that holds marketing scripts until consent says yes, stops them on opt-out and keeps first-touch campaign data. No dependencies, with tests and a live demo.

JavaScript
node:test
GitHub Pages

View code Live demo

sitemap-coverage-audit

A command-line tool that compares a sitemap with a monitoring list and shows the pages nobody is watching. Comes with a Claude Code skill so an agent can run it and open a pull request for review.

Python 3.12
pytest
uv
Claude Code

View code

About

Where I have worked

2024 – 2026
Software Engineer, Flex
MarTech & Growth Engineering · Remote
US fintech that lets renters split their rent into two payments. I owned the marketing web platform end to end (getflex.com, paid landing pages, help centers, tracking and consent, monitoring) and the AI tooling around it.
2018 – 2024
WordPress & WooCommerce developer
Tuyo Tienda
Delivered business sites and online stores for small-business clients: Elementor Pro and Divi builds, WooCommerce with MercadoPago checkout and a Tutor LMS course platform, plus custom plugins (a WooCommerce AJAX category filter), VPS hosting and backups, and n8n automations.

Tools I use

Analytics & consent

Amplitude (Analytics, Experiment, Session Replay)
GA4
Google Tag Manager
Transcend Airgap (CPRA)
LaunchDarkly
A/B test instrumentation

Web & content

Webflow (CMS, Data API)
Unbounce
Zendesk Guide
WordPress/WooCommerce
HubSpot
SEO & JSON-LD

Data, cloud & code

SQL
Snowflake
Athena
JavaScript
TypeScript
Python
AWS (Lambda, S3, CloudFront)
Terraform
Datadog
PagerDuty
GitHub Actions

Ad, lifecycle & AI

Google Ads
Meta
TikTok
Singular
Braze
Iterable
Claude Code (plugins, skills, agents)
MCP
Cursor

Languages: Spanish (native) · English (professional, used daily with a US team)

Contact

Want to talk about growth engineering, tracking or AI automation?

I am open to remote roles. Write to me.

agugentile@gmail.com