What we're learning. What our customers are teaching us. Occasionally, what we got wrong. Long-form only. No listicles.
Eighteen months on the new runtime. 13x p99 latency win, 50x concurrent run ceiling. Here's the 32,000-foot summary of the architecture, and the two things we'd do differently next time.
We spent four months making SQL optional. We also spent four months making sure it was never invisible.
You're paying by the webhook, the contact, the field, the report, and the seat. Here's the math across three big vendors.
Idempotency keys, per-tenant queues, and why we gave up on "exactly-once" and leaned into at-least-once plus dedupe.
Exports, mapping collisions, and the one custom field that took longer than the other 212 combined.
A candid look at our AI deal-summarizer experiment. The data said "adoption." The customer interviews said something else.
The spreadsheet template we share with customers on day one. Weighted probability x stage x decay, with every number traceable.
The identity-tax pricing pattern is actively hostile to security. Here's why we refused to do it.
The tooling we built, the one mistake we made on region 7, and why you should use logical replication even if you don't think you need it.
Acceptance rates, completion rates, quality of signal, and what it costs us per hire. The conclusion: we're not going back.