What a small team can ship once the machine takes the repetitive part
We are not selling "we use AI": we show you the pipeline and the numbers from our own site. People deciding, agents executing, and a CI that refuses to let half-finished work through.
Soamee's methodology combines the CRAFT process (Consult, Research, Architect, Forge, Track) with AI-agent-assisted development and a continuous integration setup that blocks any incomplete delivery. Product, architecture and design decisions are always human; agents handle the repetitive, verifiable work: boilerplate, translations, migrations, tests and SEO checks. Every change goes through human review and 22 CI automations that verify translation coverage, images, SEO metadata, Schema.org, social cards and broken links before reaching production. The measurable result on our own repository: 996 pages published in five languages, 635 articles and 130 case studies kept at parity by a boutique team based in Madrid.
CRAFT, with an agent layer underneath
The process does not change: what changes is who does each task inside each phase. Above, the quality gates nothing skips. Below, what we delegate to agents.
Nothing reaches production without clearing all five gates, no matter who wrote it.
Human review · Tests · CI · SEO audit · Deploy
We learn the business and the problem. We leave with measurable goals and a preliminary scope.
We analyse users, workflows and existing systems. We decide what to build and what to integrate.
We design the architecture, the data model and a delivery plan prioritised by business value.
Two-to-three-week sprints with a demo at the end of each one. Working software from month one.
We deploy, measure real usage and evolve the system: maintenance, improvements and new modules.
Nothing reaches production without clearing all five gates, no matter who wrote it.
We learn the business and the problem. We leave with measurable goals and a preliminary scope.
Market and competitor research, session transcription and summaries.
We analyse users, workflows and existing systems. We decide what to build and what to integrate.
Reading third-party APIs and docs, throwaway prototypes to validate integrations.
We design the architecture, the data model and a delivery plan prioritised by business value.
Project scaffolding, schemas, typings and API contracts from the human decision.
Two-to-three-week sprints with a demo at the end of each one. Working software from month one.
Boilerplate, tests, mechanical refactors, bulk migrations and translations.
We deploy, measure real usage and evolve the system: maintenance, improvements and new modules.
Metric watching, regression detection and report drafts.
What a person decides and what the machine accelerates
Every client asks the honest question: "if AI does it, why am I paying you?". Here is the answer, unvarnished.
Always human
We delegate none of this. It is where a project is won or lost.
- What to build and in what order: scope, priorities and cuts
- Architecture, data model and the decisions that are expensive to undo
- Visual and product design — no generic templates
- Review of every line of code before it merges
- Direct contact with you: whoever writes the code is in the demo
- Accountability for the outcome. The tool signs nothing
Accelerated by agents
Verifiable, repetitive, boring work. Exactly where a machine never tires.
- Boilerplate and scaffolding from an architecture already decided
- Translations into five languages, reviewed afterwards
- Bulk migrations and mechanical refactors across hundreds of files
- Tests, fixtures and CI scripts
- Crawling third-party API documentation
- Accessibility, technical SEO and metadata checks
We put it in writing: we published a post-mortem of our own repository with what worked and what we had to redo. Read the repository post-mortem →
The gates nothing skips
Writing code fast is easy. The hard part is making sure nothing is left half-done at speed. That is why incomplete work never even reaches the main branch.
Translation coverage
If a page or an article ships in Spanish without its four translations, the push fails. No orphan content.
npm run check:translations Verified images
Checks that every referenced image exists and that two different entries do not reuse the same file.
npm run check:images SEO audit
Titles, descriptions, h1s, canonicals and hreflang for every built URL. No manual exceptions.
npm run check:seo Schema.org
Validation of each template’s JSON-LD so search engines and LLMs understand the page.
npm run check:schema Social cards
Every URL gets its own social card with its own copy, rendered on demand and verified in CI.
npm run check:og Pre-push hook
The same checks run on your machine before anything is pushed. CI is the safety net, not the first warning.
.githooks/pre-push On top of that run the scheduled audits: Lighthouse, broken links, metadata guard and ranking snapshots. 22 automations in total.
Our own site is the case study
We cannot show you our clients’ internal numbers. We can open up the soamee.com repository, which is built with exactly the same pipeline.
Published and maintained, dynamic routes excluded
26 cases × 5 languages, always in sync
127 Spanish originals plus their 4 translations
Our own design system, no bought templates
Published pages per language
Multilingual parity is the most honest metric of a pipeline: it is the first thing any capacity-constrained team abandons.
What changes in your project
Three concrete consequences of working this way, with no promises we cannot keep.
More surface per sprint
What another agency parks in "phase 2" — the other languages, the admin panel, the documentation — usually fits inside the initial scope.
Less hidden debt
Automated checks catch daily what a human team only finds in next year’s audit.
No black box
We tell you which part an agent generated and which part a person reviewed. And the code is yours from the first commit.
Frequently asked questions
The ones we get on the first call, answered before you have to ask.
Does AI write all the code in my project?
If you use AI, is the code lower quality?
Who owns the code?
Can I see which part an agent generated?
Does this methodology suit every project?
What about my confidential data?
Tell us your challenge. We'll propose a solution.
No commitment. Within 24 hours, you'll receive a proposal with scope, timeline and budget. No fine print.