Eight stages between your prompt and a working app.
VULK isn't a single LLM call. It's a deterministic pipeline that classifies, plans, routes, generates, validates, repairs, previews and deploys — with the right model picked for each stage and an autofixer loop when a step fails.
- 01
Fast classifier
Classify the prompt
What surface does the user want? Web app, mobile (Flutter), Shopify theme, marketing site, internal tool? A small, fast classifier picks the right pipeline before generation starts — wrong surface is the most expensive failure mode.
- 02
Heavy planner (Claude Opus / Gemini Pro)
Plan the architecture
A planner expands the prompt into a feature list, schema sketch, route map, model decisions and acceptance criteria. The plan is what the user sees scrolling first; nothing is generated before the plan is locked.
- 03
Router (deterministic)
Route models per phase
Each subsequent step gets the model that performs best for it on our internal benchmarks: a fast model writes scaffolding, a heavy model writes business logic, a specialised model writes 3D / Flutter / Liquid. We change the routing when a better model lands.
- 04
Generator (per-phase model)
Generate the code
Every file is emitted as a structured XML action — VULK never just streams freeform code. The XML lets us track which files exist, validate before write, and surgically edit later without regenerating the whole project.
- 05
Validator (rules + small model)
Validate against the build
Type-check, dependency resolve, vite build, schema sanity. Static analysis runs synchronously — we won't show you a preview that doesn't compile. Failures route into the autofixer; successes go straight to preview.
- 06
Repairer (heavy model)
Autofix the failures
When the validator finds an error, the autofixer reads the actual error output and surgically edits the failing files. It loops until green or until a hard cap — and reports honestly when it can't fix something instead of pretending it did.
- 07
Boot the microVM preview
A Firecracker microVM provisions in seconds: kernel, rootfs, dependencies, the project. Live preview is real, not a sandboxed iframe — your generated app runs the same way it will in production.
- 08
Deploy where it lives
One click ships to Cloudflare Pages, Vercel, or your own infra. Mobile builds (Flutter) hand off to EAS / native pipelines with the right metadata for App Store + Play submission.
Which model runs which stage.
Routing is deterministic and changes when benchmarks change. Specific model identifiers shift weekly — the policy below is stable.
| Stage | Model class | Why |
|---|---|---|
| Classification | Small / fast | Cheap, sub-second |
| Planning | Heavy reasoning | Single shot, longest context |
| Generation (web) | Heavy code-tuned | Tightest output rules |
| Generation (mobile) | Specialised Flutter-aware | Widget-tree correctness > brevity |
| Generation (3D) | Specialised Three.js | Demands correct geometry / shaders |
| Edit (surgical) | Heavy code-tuned | Smaller diff window, must keep file alive |
| Repair | Heavy reasoning | Reads compile errors as input |
| Summarise | Small / fast | Plan recap, completion notes |
What happens when generation goes wrong.
Most AI app builders stream code, hand it back, and hope. VULK assumes the first generation will fail at least once and treats recovery as a first-class step.
- Static checks first. TypeScript, eslint-blocking rules, dependency resolution. Cheaper than running the build.
- Build the project. Vite for web, Flutter analyzer for mobile, theme-check for Liquid. The real build, not a simulation.
- Capture the error verbatim. The fixer reads the actual compiler output, not a paraphrased summary.
- Surgical edit. The repair model emits a vulkEdit action targeting only the failing files — full regeneration is the exception, not the rule.
- Loop until green or cap. Maximum three repair passes. After that we surface the failure honestly with the error and a one-click "regenerate from scratch".
Want to feel it?
The fastest way to understand the pipeline is to watch it run. Start on a paid plan — full plan credits, cancel anytime.
Last reviewed: April 30, 2026