The problem
Migrating project data between Azure DevOps and Jira is slow, manual, and error-prone. Enterprise teams lose ticket history, break parent-child relationships, and spend weeks on one-off scripts that nobody wants to maintain. Consultants charge six figures; internal teams stall.
Migrayt exists to make this self-serve: connect your source, map your fields, preview a dry run, then execute a checkpointed bulk write, with a clear audit trail of what ran and when.
Architecture
The platform is split into three isolated stages: intake & catalog (secure connect, full project scan, pricing from real volume), mapping intelligence (types, fields, sprints and components; AI drafts mappings, humans approve), and execution & delivery (dry-run preview, checkout, checkpointed bulk writes). Each stage is separate so customers always know what ran and when.
Infrastructure runs on AWS (ECS for compute, RDS for state, SQS for job queues, VPC for tenant isolation), all provisioned with Terraform. Stripe handles usage-based billing. Azure DevOps CI/CD pipelines deploy the application layer.
AI with Zero Data Retention
Field mapping is where AI earns its place. Source and destination schemas rarely align one-to-one; an LLM can propose sensible mappings from context, but enterprise buyers require strict data handling. Migrayt's AI sanitisation layer processes data under a Zero Data Retention policy: no customer ticket content is stored after the mapping step completes.
That constraint shaped the architecture: ephemeral processing, explicit consent gates, and audit logs that record *what* was mapped without persisting *what* was in the tickets.
What I'd do differently
Ship the dry-run preview earlier. Customers needed to *see* the migration output before paying; once we made that the centrepiece of the funnel, conversion improved. The technical work was always there; the product framing caught up later.
More from the build log
Aug 2026 · 7 min
I expected the model to believe old documents. It did something else.
Building a conflicting-evidence benchmark, breaking the harness twice, and watching the main hypothesis fail.
2026 · 14 min
Building a draggable card-canvas hero (with the code)
A breakdown of this site's hero: ten draggable cards, centred copy, mono hover hints, and copy-paste code for each one.