Roberto Galán, Senior AI Engineer

Every change to 16 internal apps queued behind one engineer. Four departments waited about a week for small fixes.

Teams ask for a change in plain language; agents build it, review it, deploy it and test it in the live app. Multi-agent pipeline on Claude Code: router, builder in an isolated branch, integration and security reviewers with veto, health-checked deploys with auto-rollback, QA agent on the live app.

Four departments ship their own changes in under an hour, with no new engineering hires. Broken releases are stopped before users see them, failed deploys roll back on their own, and risky changes wait for a person to approve.

Lead time from request to production

Every video ad variant needed an editor: scripting, voice, music, captions and cuts done by hand, capping output per week.

Raw footage goes in, finished ads come out: formats, captions, music and cuts done automatically and checked before delivery. Vision models, generative fill, WhisperX, FFmpeg, automated QA on every render.

Production per ad went from about 4 hours to under 8 minutes. The ops team runs it, and editors moved from assembly to review.

Shorter cycle per ad · 35,000 ads delivered

Generic AI ad copy ignored who the customers actually are, and every new client took two weeks of research before the first concept.

Takes a brand's best-performing ad, works out why it wins, and writes new versions for real customer profiles. A brand gate checks logo, colours, claims and legibility before anything ships. Retrieval over 13k+ enriched customer profiles, multimodal LLMs, image generation, automated brand gate, Meta Marketing API.

One winning ad becomes 124 on-brand variations, and client onboarding dropped from two weeks to two days.

On-brand variations from one winning ad

Customers order through WhatsApp voice notes, photos of handwritten lists and PDFs. Staff retyped every line into the SAP ERP.

An agent reads each order, matches it to the customer's catalogue and prices, and asks a person only when unsure. WhatsApp Business API, structured extraction, per-line confidence scoring, human-in-the-loop, SAP export.

Order entry became exception handling: the team reviews the 0.2%, not the whole queue, and the same headcount handles more volume.

Of sales orders processed straight through into SAP