How we use AI

Not "ask an LLM to rewrite it." Deterministic parsing, AI-accelerated generation, human-approved judgment.

"AI-powered migration" means very different things depending on who's selling it. Some of it is a chatbot given your source files and asked to translate them, with no structural guarantee it preserved your business logic. That's not what we do, and this page explains the actual difference.

Draft. Pending Pedro's review before publication — describes the method at the level already implied by the live wedge pages' "How It Works" sections; nothing new is claimed here.

The actual pipeline

Three stages. Each one inspectable, not a black box.

Deterministic parsing

We parse your source estate — OutSystems OML, Oracle Forms FMB/PL-SQL, or whatever your stack's actual file format is — into a complete structural map: every entity, screen, business rule, integration, and data flow. This step is deterministic parsing against the platform's real format, not an AI guessing at what your code probably does.

AI-accelerated generation

From that structural map, AI generates idiomatic, maintainable code in your target stack at volume — this is where the speed comes from. Every structural and architectural decision is reviewed by a senior architect before it's accepted, not auto-approved because the AI produced it.

Behavioral equivalence testing

Before any cutover, AI synthesizes tests that compare the new system's output against the live legacy system's output on the same inputs. You migrate on a passing diff, not on faith that the conversion "looks right."

Why this distinction actually matters

Two different claims get called "AI migration."

"Ask the model to rewrite it"

Point a general-purpose LLM at source files and ask for a translation. Fast to demo, but there's no structural guarantee it preserved every business rule, every edge case, every integration quirk — the model can produce confident, plausible-looking code that quietly drops behavior. You often don't find out until production.

Parse, generate, verify

Start from a structural map extracted deterministically from your actual source format — not the model's interpretation of what the code "probably" does. Generation is AI-accelerated for speed. Every result is checked against the live system's real behavior before it ships. Confidence comes from the test, not from the demo.

Who owns what

AI does the volume. Senior architects own the judgment.

For the full breakdown of data handling, IP ownership, and the human-approval gates in the process, see Security, IP & how we use AI.

Next step

See it on your own estate, not a demo.

The Proof of Architecture Diagnostic runs the real pipeline against one of your actual modules — not a curated example.