Your engineers are about to make a decision we have already made.
Short, opinionated briefs on patterns that held up in production: what to get right, what breaks, and where AI pays off on top. Written for the engineers and architects doing the work. Free to read, no form.
A field guide to branch-per-environment CI/CD
The delivery model where merge is the only deploy verb — feature branches run Dev, main runs QA, a release branch runs Prod, and Terraform makes it true on every merge.
Most teams run two systems that disagree about what is deployed: version control, and a deployment dashboard bolted next to it. For a platform team, there is a simpler contract — bind each branch tier to exactly one environment, make every merge a Terraform apply, and route environment config fixes back down the ladder so nothing tested is ever lost. This is the model that delivered a university data platform of fifty-plus Lambda functions and twenty-plus Step Functions workflows. Four moves that make it work, the honest constraints, and where it stops scaling.
A field guide to geospatial lookup
Answering "which areas contain this point — and which points fall inside this area?" with plain PostgreSQL and PostGIS, fast enough to sit on the request path.
The moment polygons appear on a roadmap, most teams assume they need a specialty geospatial engine. In production, plain PostgreSQL with PostGIS answers point-in-polygon questions in milliseconds — if you respect how a spatial lookup executes, and if boundary data ships with the discipline of a software release: versioned, activated by a pointer flip, archived instead of deleted. Four moves that keep the hot path fast, three places the database will quietly betray you, and where AI pays off on top.
A field guide to event-driven orchestration
Retiring the orchestration server — replacing Airflow and commercial schedulers with Step Functions, EventBridge, and Lambda that only exist while a pipeline is running.
Most data platforms carry an orchestration server whose only job is to start other work — and it is usually the least reliable, most patched, most babysat box in the stack. The alternative is not a better scheduler. It is no scheduler: pipelines that assemble themselves from events, run as state machines, and cost nothing between runs. We have retired the scheduler three times now — at a research university, inside Salesforce, and at Taco Bell — and the pattern holds. Four moves that make it work, and the honest costs to weigh before you commit.
A field guide to auto-relationalization
Turning nested API exports — Workday, PeopleSoft, Salesforce — into a proper relational model, without rebuilding it by hand every release.
Most data teams treat nested-to-relational as a design problem — someone draws the target schema, someone else writes the mapping code. It is actually an extraction problem. The relational model is already in the export, encoded in the path structure of the source itself, and the work is not designing it. The work is reading it. Four moves that make the extraction reliable, three details that make it survive schema drift, and where AI starts paying off on top.
A field guide to selective promotion
How a large engineering team shares one development sandbox — and still ships any single feature to production, on any schedule.
Most release models quietly force a choice: integrate and test everything together and therefore ship it together, or ship features independently and never test them in the combination they will actually run in. There is a third option large engineering organizations eventually need — a shared sandbox for integrated testing plus selective per-feature promotion. Here is what makes it work, and the edge cases to design for deliberately.
A field guide to the lean data lake
Assembling S3, Lambda, EMR, and Iceberg into a production-grade lake — and the four things you have to get right.
Many teams reach for a managed lakehouse because they assume the alternative is operationally too heavy. In our experience running production lakes at scale, the opposite is true. A small set of primitives, assembled with the right discipline, produces a leaner, cheaper, more debuggable lake than any managed platform — and it leaves every AI exit ramp wide open.
One email per new brief.
Get the next brief by email.
A short note when a new field brief publishes: the point, and the link. One-click unsubscribe.
You get one confirmation email first. Nothing sends until you click it.
Bring us the problem you can already name.
The first conversation is free. If we are the right firm for it, we propose a scope in writing. If we are not, we tell you who to call.