Sessions

Latticework runs ten thousand builds a day and nobody watches them. Getting there took three years, one deleted queueing system, and a rule we now apply to everything: if a failure needs a human to interpret it, it is not finished. This keynote is the rule, the three times it was wrong, and what a boring pipeline costs to keep boring.

Vela migrated four hundred services onto a new platform over eighteen months without a freeze and without a migration team. The trick was making the new path cheaper than the old one for the person doing the work that week, and never once asking a team to migrate as a favour.

A year of code generation confined to the parts of the codebase nobody enjoys: adapters, fixtures, migrations, tests for pure functions. What it was good at, where the review burden ate the gain, and the one category of change we stopped allowing it near.

Cobalt Grid's telemetry cost more than the compute it was watching. This is the audit that found it, the four cardinality decisions responsible for most of it, and the sampling policy that cut the bill by seventy per cent without losing a single incident's worth of detail.

Compiling our compiler to WebAssembly turned a twelve-second round trip into an editor that answers as you type. The talk covers what had to change in the compiler, what WebAssembly still makes hard, and how we kept one implementation rather than acquiring a second one.

Alderwood never split the monolith, and it is the most profitable system the company has. A keynote about the decisions that kept it habitable — module boundaries enforced in CI, a test suite treated as a product, and a standing refusal to adopt anything before the second team asks for it.

Substrate measured the gap between a new hire's start date and their first merged change, then spent a quarter closing it. Most of the win was not tooling: it was deleting the seventeen documents that disagreed with each other about how to run the app locally.

Three projects where fine-tuning was the obvious move and the wrong one, with the numbers that showed it. Retrieval beat it, then prompt structure beat it, then a smaller model with a better evaluation set beat it. The fourth project fine-tuned, and this is how we knew that one was different.

Tidewater ships daily from a shared trunk with no release manager and no cut meeting. The talk walks through the automation that replaced both, what happens when the train has to stop, and why the rollback path gets exercised on purpose every week.

What actually breaks in a monorepo at forty teams: ownership, review latency, and a CI graph nobody can hold in their head. Vantage Foundry's answers to each, including the one policy — no cross-team import without a declared interface — that did more than all the tooling combined.

Every model in production is a thin layer over a data platform somebody has to keep correct. A keynote about the unglamorous half: lineage that survives a schema change, backfills that do not lie, and why the freshness guarantee is the one number the model team should be reading.

Quanta went two years without a severity-one incident, and it was not because the estate got simpler. A talk about the four habits that did it — fewer knobs, ruthless defaults, a change calendar with teeth, and treating near-misses as the real signal — with the numbers behind each.

In clinical summarisation there is no gold answer to score against, and accuracy is the wrong question anyway. How Kestrel Health built an evaluation set clinicians agreed with, what inter-rater disagreement told us about the task, and the failure modes that only ever showed up in review.

Meridian Freight's retrieval runs on the search index it already had. The talk compares it against the vector database it replaced on the same traffic — recall, latency, cost and the operational surface — and is honest about the two query shapes where the vector store still wins.

Finch House deleted two thirds of its engineering metrics and got better at the job. Which ones went, what replaced them, and how to tell the difference between a measure that informs a decision and one that exists because it was easy to collect.