GridWorks was founded in 2019 by three engineers who had spent a decade being the people called at 3am. We stayed deliberately small: everyone who works on your systems has operated systems at scale themselves, and the person in the sales conversation is the person who will be in your repositories.
These are not values on a wall. They are the positions we take in design reviews, and the reason some engagements do not happen.
Every novel component in your stack is an on-call burden and a hiring constraint. We spend innovation tokens where they create differentiation for your product, and nowhere else. Postgres until Postgres genuinely stops working.
We do not accept "it feels faster" as an outcome. Every engagement opens by establishing a baseline — latency histograms, change failure rate, cost per request — and closes by comparing against it. Including when the comparison is unflattering.
Chasing five nines on a service where users would not notice four is an expensive way to slow your roadmap down. We help you pick a target, defend it, and spend the remaining budget on shipping.
A golden path that engineers route around is a failed platform, no matter how elegant the abstraction. We measure adoption, we ask why people opt out, and we treat the developer as the customer.
Anyone can automate the happy path. The value is in the rollback that runs without a human, the failover that has been drilled this quarter, and the restore that someone has actually tested against production-sized data.
We build on upstream open source, in your accounts, in your repositories. There is no GridWorks agent, no proprietary controller, no licence. If we are still essential after twelve months, we have done the job badly.
Nobody has ever been fired at a GridWorks post-incident review, because "human error" is the beginning of an investigation, not the end of one. If a single mistake could take production down, the system permitted it.
We have turned down work where the real problem was organisational, where the timeline was fiction, or where the client wanted a rubber stamp on a decision already made. We will tell you that in the first call rather than the third invoice.
Six of the fourteen; the rest are on client engagements and prefer not to be listed. Every principal has run production at a scale where the failure modes get interesting.
Dr. Anja Vosloo
Spent six years on edge infrastructure at a global CDN, where she led the migration of the anycast routing layer off a bespoke control plane. PhD in distributed consensus; still maintains a Raft implementation nobody asked for.
Marcus Ilori
Nine years of SRE across a hyperscaler and a UK challenger bank, the second of which taught him that regulators and error budgets are compatible if you write things down. Wrote the incident command handbook we give clients.
Priya Raghunathan
Built the deployment platform at a commerce company through the years it went from 80 to 900 engineers. Has strong opinions about merge queues and a documented grudge against long-lived feature branches.
Tomás Ferreira
Former capacity planner on a public cloud compute team, so he knows what the pricing page does not tell you. Built the unit-economics model we use, and once found a $340k/year saving in a single misconfigured storage lifecycle rule.
Ines Marchetti
OpenTelemetry collector contributor and maintainer of two receivers. Came from an APM vendor's backend team, which is why she is unusually direct about what telemetry actually costs to store and query.
Kwame Osei
Worked on artifact signing and provenance tooling before it had a marketing category. Leads our SLSA hardening work and reviews every policy pack before it reaches a client cluster.
◆ Paired working session · half a day, paid
No take-home exercises and no algorithm rounds. Candidates walk us through a system they have operated and an incident they got wrong, and then pair with two of us on a real client-shaped problem for half a day, paid. We have made eleven hires this way and lost one.
Every engineer gets one day a week for upstream work. It is not altruism — patching a bug in the collector is cheaper than maintaining a fork of it across nine clients.
| Project | Area | Contribution | Status |
|---|---|---|---|
| opentelemetry-collector-contrib | Telemetry | Two receivers maintained, 340+ commitsplus the tail-sampling memory regression fix in v0.98 | maintainer |
| karpenter | Compute | Consolidation edge cases with PDB interaction6 merged PRs, 2 open RFCs | contributor |
| argo-rollouts | Delivery | Prometheus analysis provider — exemplar supportand a fix for aborted-rollout metric leakage | contributor |
| crossplane-contrib/provider-aws | Substrate | RDS composition defaults, PITR and CMK wiringreduced a 200-line claim to 12 | contributor |
| kyverno | Policy | Image verification policy library for cosign keylessnow part of the upstream sample set | contributor |
| gw/slo-generator | Our own | Generates multi-window burn-rate rules from a service manifestApache 2.0, 2.1k stars | maintainer |
| gw/tf-module-library | Our own | The 41 Terraform modules we hand to clients, publishedApache 2.0, terratest coverage 84% | maintainer |
commits merged 412 projects touched 19 issues triaged 288 CVEs reported 3 talks given 7 ────────────────────────── maintainer seats 3 forks maintained 0 ↑ this is the metric we actually care about
The first conversation is with whoever would lead your engagement. If we are not the right fit we will say so on that call and, where we can, point you at someone who is.