SovrenAI
Anyone describes a website in plain English and gets a deployed multipage site
We engineer Atomic Builder, SovrenAI's AI website builder: a prompt runs through a pipeline of AI skills, a validation engine checks the generated code and an auto-fix step repairs it. Users edit with an agent, preview live and deploy on their own database.
B2B SoftwareAtomic Builder, an AI website builderBuilt from zero
- Generates
- Multipage sites from one plain English prompt
- Checks
- Validation engine and auto-fix on generated code
- Runs on
- The user's own Supabase, connected per project
- Billing
- Credits, subscription plans and credit packs
The challenge
Generating code from a prompt is easy to demo. Getting a site that builds every time, edits without breaking and runs on the user's own data is the product.
What we did
We built generation as a pipeline of AI skills grounded in a reusable template, with a validation engine, auto-fix, an edit agent, live preview, credit billing and push to GitHub.
The outcome
The builder runs in separate Dev and Prod environments, its SEO is live on its own domain, and every feature is in final QA before the full release.
The starting point
Atomic Builder, SovrenAI's AI website builder, lets a person describe a website in plain English and get a deployed multipage site, with live preview, edits by an agent and push to GitHub. Generation is paid for in credits.
What had to change
A generator that works in a demo still fails on real prompts: code that does not build, slow runs, lost progress and invented packages. The builder needed generation that checks and repairs itself, and edits that leave a working site.
What could not break
- Generated code passes validation before anyone sees it
- Users connect their own Supabase instance, remembered per project
- Only real, reputable npm packages make it into a site
- A long generation survives a pause or a lost connection
- Builds and deployments follow the user's plan
The calls we made
We framed the builder as a production pipeline that checks its own output, not a single prompt that writes a site.
- 01
Skills, then validation, then auto-fix
Generation runs as modular AI skills, a skill engine with producer skills, a guidance skill and a content producer node, and every result passes a validation engine with an auto-fix framework behind it.
Ruled outOne long prompt that writes the whole site unchecked.
- 02
Ground generation in a real template
Sites are built from a reusable boilerplate of 52 files, 11 page sections, 15 UI components and 7 starter kits, plus a golden template and TypeScript auto-fixes, so output starts from code that already works.
- 03
Plan the change, then edit
An LLM change planner decides which files a request touches before the edit agent applies it, and a build validator checks the result.
Ruled outThe regex file picker that used to choose which files an edit would change.
- 04
Check packages against the live registry
A live npm registry check with reputation gating decides which dependencies a generated site may use, covered by 34 automated tests.
Ruled outHardcoded package checks, which missed invented dependencies.
- 05
The user's database, not the builder's
Users connect their own Supabase instance through OAuth, remembered per project, so a deployed site keeps its data in the user's account.
Ruled outRunning every user's site on the builder's own Supabase organization.
How the system fits together
What we built
Skill pipeline with validation
Modular AI skills generate each site from the template system, a validation engine checks the output and an auto-fix framework repairs it. Steps run in parallel with output limits and a soft timeout, and generation resumes after a pause or a lost connection.
Edit agent and live preview
An edit agent applies planned changes to a site running in a sandbox preview, with a design resolver, a smart questionnaire and internationalization in the builder.
Website clone and capture
A Playwright service captures a URL securely, processes and tiles the screenshots and reads them with AI vision, so a user can start from a site they already like.
Credits and plans
A redesigned credit system and token economy with subscription plans and credit packs, and plan based limits on builds and deployments.
Accounts, images and SEO
Email verification, Google sign in and password reset, image search on Pixabay with Openverse as backup and Azure storage, plus a sitemap, per page meta and a blog module for the builder's own site.
How it works
Describe
The user writes what the site should be, and a smart questionnaire fills in the gaps.
Generate
AI skills build every page from the template system, running steps in parallel.
Validate and fix
The validation engine checks the code, auto-fix repairs it and the npm check rejects unknown packages.
Preview and edit
The site runs in a sandbox preview, and the edit agent applies each planned change.
Deploy
The user connects their own Supabase and the site is pushed to GitHub and deployed within their plan.
A closer look
What changed
The builder runs in separate Dev and Prod environments, and its SEO is live on the production domain. Generation is faster, error free in internal QA and resumable, and every feature is going through internal QA before the full release.
Under the hood
- Datics role
- Product built by Datics
- Integrations
- Supabase through OAuth, Stripe billing, GitHub, Pixabay and Openverse images with Azure storage, Azure Communication Services for email, Google sign in and Playwright for capture.
- Data boundary
- Each deployed site connects the user's own Supabase instance, remembered per project, rather than the builder's organization.
- Supply chain
- Dependencies are checked against the live npm registry with reputation gating before a generated site can use them.
- Deployment
- Separate Dev and Prod environments, sandboxes for live preview with idle reaping, and Microsoft Clarity on production.
AI and models
Data and retrieval
Cloud and infrastructure
Apps and APIs
Also in the stack
How it shipped
Engine and templates
The AI skill pipeline, the template system, design and UX work, preview fixes, security headers and login rate limiting, with QA every day.
Clone, credits and environments
The website clone and capture feature, the redesigned credit system, and separate Dev and Prod environments with the builder deployed.
Own database and SEO
Users connect their own Supabase, the npm dependency check replaces hardcoded rules, plan based deployment limits, and SEO goes live on the builder's domain.
Faster, resumable generation
Parallel steps, runtime validation and the golden template made generation faster and error free, generation became resumable, and new image providers and the blog module shipped.
Final QA
Internal QA of every completed feature before the full release.
In the founder's repositories
Our engineers work in the founder's own repositories and own deployment, alongside the changes he pushes himself.
Weekly sprints, daily QA
Each week ends with a progress report to the founder, and QA tests generation, preview and deployment every day, retesting each fix.
Measure, then speed up
We timed generation end to end, then ran steps in parallel, added output limits and a soft timeout, and made long runs resumable.
What this unlocked
After final QA comes the full release. The same foundation also carries the founder's enterprise and kids editions, each told in its own case study.
Related work
Have a product with the same kind of constraints?
Tell us what already exists and what you want it to do. We will tell you what it takes.