Engineering note · Technical support engineering · by Ayman Sbeiti — I support high-trust software platforms · Hiring?
Engineering playbook · Tooling
Preparing a legacy dataset for migration with an AI-assisted script
Core principle
AI generated the scaffolding in minutes. Deciding what the data needed to look like on the other side, and checking that it did, was the actual work.
Real-world example
Migrating FLiiP's support platform meant moving years of accumulated tickets and contacts between systems with different data models. The legacy dataset had everything long-lived data accumulates: duplicates, dead records, inconsistent fields. I used Dust to help build a script that processed a JSON export of the legacy data, and the AI generated the scaffolding (parsing, iteration, the shape of the transformation) quickly and mostly correctly.
Why it happens
The AI could not know what correct meant for this dataset: which fields mattered on the receiving side, which duplicates were actually the same customer, which "empty" values were meaningful and which were noise. That judgment, and the verification that the output met it, stayed human. Cleaning before moving was non-negotiable, because migrating a mess just relocates it.
What I now check
- Define what the data must look like on the other side before generating anything.
- Treat AI-generated code the way you'd treat a competent stranger's first draft: plausible structure, unverified assumptions.
- Test every assumption the script makes about the data against the actual export before anything runs for real.
- Keep the script's steps inspectable rather than clever.
- Spot-check the output against the source until the boring cases bore you.
Production takeaway
The AI saves you the typing. It doesn't save you the thinking.