An AI marketing platform, built from zero in 28 days

A client onboards in a single session. The platform diagnoses the brand, writes a quarterly brand DNA and builds a dynamic calendar for each social network.

It generates images, carousels, copy and short videos with voice, delivered as a weekly batch where each piece is approved on its own. Every edit a client makes feeds what the system learns. Clients see reports and their pieces in a client portal, and the team runs the service from an admin console.

Forbidden-claims rules by industry and country stop a piece before it exists: 10 forbidden prompts produced 0 pieces.

Everything runs on Vertex AI: Gemini and open models, plus text-to-speech and image models, with no single vendor. Cost per approved piece decides which model each station uses.

An independent panel of AI judges blocks verifiable defects and never approves; approval stays with a person. Each judge was tested on labelled examples, cases whose right answer is known, before it went into service, and calibration tools tune it against people's ratings.

• Forbidden-claims layers per industry and country: 438 data files from primary sources
• A sourced craft knowledge base per industry, niche and country across 39 countries of the Americas
• A cost ledger for every piece and every model call

TypeScript, Next.js and React on PostgreSQL. The product splits into 16 bounded contexts and exposes 53 pages and 90 API routes. 92 decision records document the architecture.

• Brain, generation and the judges panel
• Calendar, objectives, plans and reports
• Brand, channels, accounts and conversations
• Models, costs and operations
• Mail and blog

A Cloud Run service, 9 jobs and 18 schedulers sit behind a global HTTPS load balancer with Cloud Armor WAF. Cloud SQL runs on a private IP, with KMS and Workload Identity, and a BigQuery ledger holds the billing data.

All of it is OpenTofu, covered by 43 mocked-provider tests that run without touching the cloud account.

Before code leaves the engineer's machine, a push gate checks secrets, types, lint, build, dependency rules and a dead-code budget. A release tag builds 3 images, runs the migrations, deploys by image digest, the fingerprint of the exact build, and reads back the digest that is actually live.

89 releases in 10 days, from v1.0.0 on 2026-09-26 to v1.61.0 in production on 2026-10-06.

OpenTelemetry sends traces to Google Cloud, 18 alert policies watch the service, and a ledger records the cost of every model call and every piece. Each deploy is checked against the build that was tested.

Unit and integration suites, Playwright walks with axe, keyboard, dark mode, forced colours and 400% zoom, and design-fidelity checks against the approved Figma frames. A chain of repository controls runs before code lands.

The design covers 519 screen-states, and the portal redesign was judged by an independent critic. A 56-file legal pack ships with a checker that verifies every citation.

The work ran goal by goal under an explicit engineering policy, with a very small human team directing AI engineering agents.