Skip to content

migration · architecture

Legacy-to-cloud migration intelligence

A short note on using LLMs to read an old codebase, classify its shape, and emit a defensible migration plan — not to do the migration, but to decide what the target should be.

Author
Lali Devamanthri
Published
Reading time
1 min read

Placeholder article. Real content to follow. The structure below is real — it's the shape every migration-intelligence piece will take — but the words are scaffolding.

Most cloud migrations burn effort at the wrong end of the pipeline. Teams deploy a lift-and-shift machine before they've answered the question the migration exists to answer: what should the target look like? Not the servers — the shape. Services, boundaries, data ownership, failure domains.

LLMs are unreasonably good at reading a legacy codebase and proposing a cloud shape, if you constrain the task. The pipeline below is the pattern I've used across three regulated migrations.

A three-stage pipeline. A legacy monolith feeds a structured ingest layer, which forwards to an LLM reasoning stage; the reasoning stage produces a target architecture and runbook for the cloud stage to execute.

Three-stage migration intelligence pipeline: an ingest layer reads the legacy system, a model-reasoning layer proposes the cloud shape, and a planning layer emits runbooks.Legacymonolith · on-premSOURCEIngeststructure · classifyReasonLLM + patternsPlanrunbook · IaCCloudtarget shapeTARGET

Read the legacy. Reason about the shape. Emit a plan.

Ingest — structure first

Before any reasoning, you need the legacy code in a structured form. Classify every module: domain, dependency, IO surface. The output is a graph, not a summary.

Reason — with patterns, not vibes

The LLM's job is to map the graph onto known patterns (strangler, anti-corruption layer, read-side projection) with evidence. Every decision ends up with a citation to the legacy code it came from.

Plan — runbooks, not slide decks

The output is an incremental plan the migration team can actually execute: IaC scaffolding, service boundaries, migration waves, the data-gravity calls. Runbooks, not slide decks.

The interesting part is not the move. The interesting part is the decision about what to move into. This pipeline makes that decision defensible.

End of article

Building something AI-shaped for healthcare or fintech?

I work with a small number of teams at a time on integration architecture, eval pipelines, and getting models into regulated production. If the system you're designing rhymes with the one above, let's talk.