Earlier Labs Ltda.
An API that lists every production change before an alert, in order
We sell Earlier to whoever runs platform engineering or SRE at a software company of 80 to 400 engineers and takes incident command.
At that size a change reaches production through four or more systems, each with its own audit log, so answering what changed on an incident bridge means opening four tabs.
Earlier is an API over one append-only ledger of production changes.
Read-only connectors and signed webhooks pull in every deploy, infrastructure apply, flag flip and edge rule edit. A query names a service and when symptoms began, and returns every change that touched it in the window before, newest first. We bill only when a responder names the change behind an incident.
Earlier uses models to place a change on a service and to order a window's changes.
Rules read service tags and repository paths, but a Terraform apply across three services goes to an embedding model that matches its raw target to the service catalog; we size for it to see up to half of a new account's first-month events. A ranker suggests an order from time gap and dependency distance.
Earlier's API has no cause field: a person names the change, and both models hold back when unsure.
A service match below 0.80 confidence waits in a queue, and an unsure ranker leaves the list in plain time order, marked ranked: false. Confirmed service pairs stay with their account; only numeric alert-to-change rows join our pooled ranker corpus, which grows with every closed incident.
Earlier's load spikes during a new account's first-hour backfill and when an upstream outage opens incidents at many accounts at once.
A backfill brings 90 days of history at up to 6,000 events a minute. Change events name resources and engineers, which our buyers' security teams keep from anyone else's model, so every model runs beside the ledger on cloud infrastructure in Oregon, on capacity we control.
From Q4 2026 through Q3 2027, Earlier's hardware changes once a quarter.
In Q4 2026 an open-weight embedding model runs on cloud GPU capacity we control in Oregon, with a gradient-boosted ranker on CPU. In Q1 2027 accounts with enough confirmed pairs get the embedding model fine-tuned on them; in Q2 2027 the ranker becomes a cross-encoder on the same GPUs; in Q3 2027 nightly re-embedding moves to interruptible GPU capacity in Oregon.
Earlier Labs Ltda. was founded in Florianópolis, Brazil, in May 2026, and four months later Earlier is at the start of its hardware plan.
Keys are issued at sign-up, with no contract or card, and the first three named changes are free. The shares and bursts here are what we size for, not what we have measured.
The founders
Ximena Orozco
Founder
Took incident command at a payments company for four years and kept a tally of bridges that opened with 'did anyone deploy?'. The answer was usually sitting in an audit log nobody on the call had open.
Gustavo Tavares
CTO
Kept the Jenkins and Terraform pipelines running at a logistics software company and wrote its #deploys bot, which a third of the pipelines never posted to. He wanted the record pulled from the source, not volunteered.
Beatriz Moura
CPO
Ran feature-flag rollouts for a Seattle retailer's checkout team. A flag she turned on caused a slow hour, and nobody asked her, because flags were not deploys. She has wanted them on one list ever since.