Yann Pires-Cabral
Yann Pires-Cabral

AI Implementation Consultant & Builder

Yann Pires-Cabral

I design AI workflows, build them in Python with tests and human approval gates, and run them in live businesses.

A framed print of a painted Fenway Park leaning against a wall in daylight, styled as a room mockupTwo models wearing NekoNoir oversized streetwear on a studio backdrop: a blue tee and shorts set with a large OKARUN print and Japanese type, and an olive set with an embroidered chest motifYouTube Studio analytics for the channel, reading 2,453,983 views so far, with 130.3K hours of watch time and 3.6K subscribers gained

A music channel at 2.45M views

What I do

AI & Automation

I build agent pipelines that do the volume work, with a person at the approval gate.

  • Agentic workflows in Claude Code
  • autonomous systems design
  • pipeline architecture
  • prompt engineering
  • social publishing automation

How work gets checked before it ships

  1. Generate

    Agents research, write, render and compose the work at volume.

  2. Check

    Automated tests and gates fail the build on a known bad result before anything moves on.

  3. Approve

    A person reviews what customers will see. Publishing is a human decision.

  4. Confirm

    After publishing, the live result is read back through the platform's API to prove it landed.

WholesomeCavern

2026

332 city prints live across 26 cities, each researched, generated and composed by an agent pipeline I built, then reviewed and published by me. Every landmark is rendered in one consistent painterly style and delivered at nine print sizes with framed lifestyle mockups; the same pipeline also produces bundle, art and mockup packs. It runs behind 697 automated tests. The consistency across that volume is the argument: this is automation producing live inventory, not demos.

A framed print of a painted Fenway Park leaning against a wall in daylight, styled as a room mockup
A painted rendering of Cloud Gate in Chicago, the city skyline curving across its mirrored surface
A painted rendering of Fenway Park from the stands, the outfield and scoreboard lit under an open sky

An automated video pipeline

2025 to present

Episodes are assembled by an agent rather than by hand: it takes word-level timings straight from the voice synthesiser, locks the shot list to them, and drives HyperFrames (the third-party HTML rendering engine it is built on) to a deterministic render, so captions, charts and money figures land on the syllable rather than near it. It now produces Ritmo Global: four long-form episodes of about eleven minutes each, three of them published.

A rendered chart from the Shakira episode: where a tour's revenue came from, by leg, in millions of dollars
A money card from the Carín León episode citing a Billboard Boxscore tour gross and ticket count
A comparison chart from the same episode: gross per night at the Sphere in Las Vegas, with each figure's source named

How an episode gets made

  1. A news feed is pulled and scored, and stories too thin to carry a full episode are rejected. I pick from the shortlist.

    In a no-code builderA schedule trigger and an RSS read, then an LLM step that scores each item.

  2. Research agents pull figures, dates and quotes from named sources such as Billboard, Pollstar, Wikidata and official accounts into one brief. Every figure in the script has to trace back to a row there.

    In a no-code builderRetrieval over a knowledge base: knowledge sources in Copilot Studio, or an n8n AI Agent with a vector store tool.

  3. An agent writes eight beats, about eleven minutes of narration, in a fixed house voice, using only what the brief supports.

    In a no-code builderAn LLM node with a system prompt, the brief passed in as context.

  4. Two independent model reviews run in parallel: the script against its sources, and the footage against the script. Every finding is verified before it changes anything.

    In a no-code builderTwo parallel branches into an IF node, then a human approval: send and wait in n8n, or a Slack or email approval in Make.

  5. Agents search archives, press rooms and official posts for every person and moment the script names, and log each file's credit.

    In a no-code builderHTTP Request nodes against archive and platform APIs, results written to a sheet.

  6. The script is voiced through ElevenLabs, and the synthesiser's own word timings lock every shot to the word it belongs to.

    In a no-code builderOne text-to-speech API call; the word timings come back in the same response.

  7. An HTML composition is generated from the locked shot list, checked by automated gates, then rendered deterministically with HyperFrames.

    In a no-code builderA code or container step; no-code tools hand this to a render service.

  8. I watch the full render end to end. Only then is it uploaded, and the upload is read back to confirm the AI-content label is set.

    In a no-code builderA YouTube upload node behind a manual approval, then a read-back check.

An AI consulting agency, built with AI

2026

Artful AI Visions helps small businesses put AI to work, and I built the whole platform behind it with AI coding tools in Claude Code: I write the spec and the plan, the model writes most of the code, and the tests decide what ships. A Python program runs a lead agent and fourteen agent teams on each client and prints their roadmap as a PDF, behind 269 tests. The web backend takes orders, client accounts with Google sign-in, Stripe payments through signed webhooks and a lead system that sends and reads email through the Gmail API, behind 524 tests. I ran a security audit on it and fixed its five most serious leads ahead of launch. The site is still private, so it is shown here rather than linked.

The Artful AI Visions home page: a central Boss node wired to fourteen named AI teams around a free AI audit asking how customers buy
Seven steps from sign-up to the call: sign up, pick a plan, tell us, the teams work, checked by a person, your roadmap, we call you
The teams section of the site: each AI team on its own card with the specialist agents inside it

How a client's roadmap gets made

  1. A client picks a plan on the site and pays through Stripe; a signed webhook confirms the payment before anything starts. They fill in one form, in English, French or Portuguese.

    In a no-code builderA webhook trigger, then a form submission step.

  2. An intake agent reads the form and the client's public website, then writes a profile and a brand kit that every team is handed.

    In a no-code builderAn LLM step with the form and a web fetch as its inputs.

  3. Up to fourteen agent teams each study one part of the business, from search and social media to automation. Every reply is checked against a fixed format before the program keeps it.

    In a no-code builderParallel branches, each an AI Agent node, with a schema check on its output.

  4. A review team opens every source and checks every figure before a finding reaches me.

    In a no-code builderAn LLM judge step, then an IF node.

  5. Each finding gets Approve, Edit or Reject on a page in my browser. Nothing goes to a client without that.

    In a no-code builderA human approval step: send and wait in n8n.

  6. A lead agent writes the roadmap from the approved findings only, with a monthly cost and a setup time on every AI move. A checker then reads it against the findings once more.

    In a no-code builderAn LLM node fed the approved items, then a second LLM judge.

  7. The roadmap is printed to PDF through headless Chrome and every model call's cost is logged for the job. I upload the PDF in my dashboard and the site emails it to the client.

    In a no-code builderA PDF service step, then an email step.

  8. On the bigger plans I walk the client through it on a call, with a sheet the program wrote for me: questions to open each point and what the firm can set up.

    In a no-code builderA booking trigger from a scheduling tool.

2026

A track-to-upload video engine

A pipeline that turns a finished track into a scheduled video for a music channel in a French-language niche. It generates five artwork candidates and I pick one; from there composition, rendering, upload and scheduling run on their own, and I review each build before it goes out. The channel is not linked here: the engineering is the interesting part.

Described, not linked

A source video frame before captioning, used as the reference for alignment

2026

Word-synced lyric captions

A reusable pipeline that aligns a lyric sheet against speech-recognition timings and fills each word as it is sung. Seven songs captioned so far. Correctness here is measured, not assumed: a source claimed to be a still image is proven still by comparing pixels, because identical frame checksums can hide real motion.

Engineering

I run the production infrastructure underneath all of it, with tests that fail before I ship.

  • Python
  • WordPress and WooCommerce
  • PHP plugins
  • Playwright end-to-end testing
  • technical SEO
  • structured data
  • REST integrations

What broke, and how I found it

  1. The host said the store had outgrown its plan.

    Disk was nowhere near full. Five cleanup snippets meant to switch themselves off named a database column that does not exist, so they never did, and purged the page cache on every request for four months. Switching them off took a repeat page load from 9.16s to 0.56s.

  2. A print passed review with a fake signature on it.

    The image model sometimes paints an artist's signature into the work. One passed the thumbnail review and was caught only by a full-resolution sweep; another sat along the bottom edge, well away from the corners the check was looking at. The sweep now covers the whole bottom strip at full resolution.

  3. A stale figure sat on this page for three weeks.

    The check that polices numbers could only see figures with separators or K and M suffixes, so an outdated listing count passed with every check green. I proved the blind spot on a planted value, and the check now reads every number on the page.

2026

A structured-content publishing pipeline

A system that builds print-ready books from structured content, now live as twelve digital editions in English, French and Spanish, and in print on Amazon. Each edition is typeset in HTML, rendered to PDF and gated by 1,365 passing automated tests covering page geometry, print margins, text-layer integrity and label overflow. The product is a sensitive-subject workbook, so it is described here rather than shown. Its governing rule (the book supplies the method, the reader supplies the content) is enforced by a review record pinned to a hash of the copy, so editing the text invalidates the sign-off and fails the build.

Described, not linked

From Artful Visions

Scale: owned infrastructure

How the storefront actually runs: one custom plugin owns every runtime change, so SEO, structured data and integrations ship behind tests rather than by hand.

Read it in context

Design & Brand

I take a brand from identity through to manufactured, photographed, saleable product.

  • Brand identity
  • apparel and embroidery digitisation
  • print production
  • Adobe Creative Suite
  • Canva
  • product photography
  • ten years of music production

Artful Visions

2022 to present

The parent brand. Two sales channels, four product lines, four years. It began as custom anime portraits on a marketplace, moved onto owned infrastructure, grew an apparel label with its own manufacturing, and now runs on automation I built for it.

Two models wearing NekoNoir oversized streetwear on a studio backdrop: a blue tee and shorts set with a large OKARUN print and Japanese type, and an olive set with an embroidered chest motif
The 1ArtfulVisionsCo shop page on Etsy showing 1,985 sales and its star rating, with a notice that the shop is no longer selling
The artfulvisions.co storefront homepage, showing anime posters, custom portraits and collectibles

Origin: a marketplace shop

Custom anime portraits and physical products sold on Etsy: 1,985 sales and a 4.9 / 5 average across 647 reviews over the shop's life. I closed it deliberately when the business moved onto infrastructure I owned; the page still carries the record.

Custom acrylic card, a personalised physical product
Custom portrait listing video, from commission to packed order

Scale: owned infrastructure

WordPress and WooCommerce on managed hosting, with a custom must-use plugin as the mechanism for every runtime change: technical SEO automation, structured data, cache discipline, print-on-demand fulfilment, and an end-to-end browser test suite over the checkout. When the site went down with a 503, the cause was a PHP memory limit set above the host's per-process cap, so the interpreter could not start at all. I isolated it by requesting a file containing no code, which separated 'engine down' from 'application broken', then control-tested a second site on the same account to prove it was account-wide.

NekoNoir: the apparel line

An anime streetwear label taken from concept to manufactured product: brand identity and care labelling, artwork digitised into embroidery machine formats (.DST, .EXP, .HUS, .VP3, .XXX, .SHV), cut vinyl transfers, thirteen collections, and three photoshoots. The part most design work never reaches is the middle: turning a drawing into a file a machine can stitch.

Embroidered garment and branded packaging, produced end to end

Range

Acrylic display cards, LED mats, posters and plates alongside the apparel, each with its own production route and its own photography.

Generative motion piece. Fan art produced for an anime merchandise brand, not original IP
Generative motion piece. Fan art produced for an anime merchandise brand, not original IP

2026

Consulting & brand strategy

Ongoing brand and go-to-market work for a local wellness business: positioning, a full visual identity, a two-lane service architecture and a phased plan for adding a second service line. Client work, so nothing from it is published here without their permission.

Described, not linked

Generative Media

I produce and publish AI-generated media at volume, on a schedule, to an audience that showed up.

  • Midjourney
  • Nano Banana Pro
  • Veo 3.1
  • Suno
  • Whisper alignment
  • HyperFrames video composition
  • ffmpeg

A music channel at 2.45M views

2025 to present

An AI-produced music channel that reached 2,453,983 lifetime views and 130.3K hours of watch time, monetised (YouTube Studio, August 2026). That catalogue has since been retired. The channel now runs Ritmo Global: long-form English episodes on the business of Latin music, researched, scripted, voiced and rendered through my agent pipeline. It has 3.63K subscribers, and the first three episodes are public.

YouTube Studio analytics for the channel, reading 2,453,983 views so far, with 130.3K hours of watch time and 3.6K subscribers gained
The channel page today: three Ritmo Global episodes on Karol G, Shakira and Carín León, each titled by the money figure it investigates
A frame from the Karol G episode: a data card reading the last tour's gross, tickets and shows, above the word-synced narration caption

How I got here

Based in Lisboa, Portugal. Open to remote roles.

  • French C1
  • English C1
  • Spanish B1
  • Portuguese B1
  1. 2022 to present

    Founder, Artful Visions. Built and ran the brand end to end: two storefronts, an apparel line with its own manufacturing, three YouTube channels, and the agent pipelines that keep them supplied.

  2. 2018 to 2022

    Founder and operations manager of an automotive dealership in Brisbane. Procurement, refurbishment and direct sale of over a hundred vehicles. That is where the commercial instincts came from.

  3. 2014 to 2018

    Electrical engineering technician across New Zealand and France, on manufacturing plants, hospitals and university facilities. Fault-finding under time pressure on systems that could not stay down.