The challenge
Circini delivers some of New Zealand's largest and most complex data-platform programs — migrating tens of millions of rows from legacy systems into Snowflake for utilities, telco, banking, and government. The slowest, most repetitive part of every migration is translating a mapping document into reliable dbt transformation logic: dozens of fields, each with its own source lineage, type rules, and edge cases — hand-written and hand-checked, one model at a time.
That work is the longest pole in the program. It's unforgiving of small errors, hard to scale against deadlines, and it ties up senior engineers in mechanical SQL when their judgment is needed on lineage and governance instead.
The CallSphere solution
CallSphere built Circini a dedicated AI data-migration agent — a ChatGPT-style workspace where engineers attach the target template, the mapping document, and the existing dbt models, then ask for the change in plain language. The agent reads the field requirements and source-to-target mappings, generates Snowflake-compatible dbt models — from scratch or incrementally — and streams its reasoning live as it works.
Every model runs through a fixed Input → Enhance → Validate → Generate flow, and a Circini subject-matter expert validates and tunes each output before it ships — nothing reaches a pipeline unreviewed. All terminology stays masked and work stays inside approved cloud boundaries, under Circini's own AI policy.
The results
On Circini's live deployment, the agent has generated 78 verified dbt models — more than 16,000 lines of migration SQL — from 94 uploaded mapping and template files, across CLM-to-AMS migration work, with a human expert validating every model before it lands.
It compresses the longest pole in a data-platform program — mapping-document-to-dbt translation — from days of hand-coding to minutes of generate-and-review. That frees Circini's consultants to focus on lineage design, data governance, and the judgment calls AI shouldn't make alone, so complex migrations move faster without trading away the accuracy their enterprise clients depend on.