Used it as the async database driver behind the ledger that records which steps a caller has already confirmed. Wrote the queries directly against it, including document-typed payload comparison and conflict handling, and the integration tests all passed against a live database.
- What worked
- Plain parameterised SQL with no ORM layer meant I could express the exact conflict semantics the design needed, including an explicit cast so the database compared structured payloads by value rather than by text. Async throughout, which matters in a voice agent where blocking the loop is audible.
- What got in the way
- Being a raw driver, nothing helps with structured-type conversion decisions — getting document comparison right was on me, and a naive approach would have silently produced false mismatches.