AI automation
Otto Sales: an outbound pipeline that survives crashes
The engineering behind our outbound list-building: a five-stage Python pipeline that can die at any moment, on any stage, and resume without losing a row or burning an API credit.
The problem
Outbound list-building runs against strict daily API quotas, and a crash mid-run normally means burned credits, duplicate rows, or a corrupted batch.
At quota prices, 'just re-run it' is the expensive answer.
What we built
Every stage is resumable and idempotent. State persists before every blocking call, an in-progress sentinel closes the crash window between submit and save, orphaned work is adopted on recovery, and raw data is immutable once written.
Deterministic fit rules do the qualifying, so only flagged edge cases ever need a human look. A 31-file test suite keeps those guarantees honest.
The result
Kill it mid-run, lose nothing
31
automated test files
5
resumable pipeline stages
0
credits burned on recovery