Every CRM has 'when field X changes, do Y' automation. The hard part nobody writes about is what happens when 50,000 records change at once with events suppressed — our answer is a change manifest that turns O(records) events into O(1).
Gmail push notifications, Outlook Graph subscriptions, and raw SES MIME all converge on one ingest function. A deep dive into the dedup cascade, threading, rate limits, and realtime fan-out that turn email into support tickets.
Editing, masking, and refreshing credentials without ever sending a secret back to the browser: sentinels, nesting-aware merges, and single-flight OAuth refresh. Part 2 of a two-part series on our connections model.
How we collapsed OAuth2, client-credentials, API keys, and multi-field secrets into one blueprint table and one instance table — and made auth application declarative data instead of per-provider code. Part 1 of a two-part series.
The client side of our realtime stack: one header that kills echo everywhere, refcounted subscriptions with useSyncExternalStore, and collaborative-ish TipTap editors without CRDTs. Part 2 of a two-part series.
How we replaced hosted Pusher with a single Rust container and built a typed room registry with ACL dispatch on top of the Pusher protocol. Part 1 of a two-part series on our realtime architecture.
The orchestration layer of our sync engine: bounded resumable slices, a three-state commit verdict, opaque cursors across six pagination styles, and why the webhook is the signal but the fetch is the truth. Part 3 of a three-part series.
How synced records stay the same record: the identity cascade, the bind table, deferred relationship resolution, and why deletes are the hardest part of sync. Part 2 of a three-part series.
How we built the sync engine that pulls Shopify orders, Stripe customers, and any REST API into our entity system: streams, recursive mapping trees, and one choke-point writer. Part 1 of a three-part series.
A generic engine is a runtime, not a product. This is how ours becomes Kopilot: page-scoped tools, prompt assembly built for caching, reference cards the model renders into the UI, and autonomous runs. The final part.
A read-only chat loop is easy. An agent that sends emails, mutates records, and can still be replayed deterministically in a test suite is where the engineering lives. Part two of three.
How we built the runtime that powers Kopilot: parts-based messages, async-generator streaming, and the safety valves that keep a model from running away. The first of a three-part deep dive.