Hybrid cloud · DevOps · AI-assisted engineering · Vilnius
Cloud, on-prem, or both. We design, migrate, and run infrastructure across AWS, GCP, Azure, VMware, and self-hosted Kubernetes — without vendor lock-in.
Infrastructure, delivery and models — across AWS, GCP, Azure, VMware and self-hosted Kubernetes. Same discipline in each column; no column is the side business.
Cloud, on-prem, or both — self-healing, and portable out of either.
A running environment you own — architecture and network design, hardened cluster, secrets and policy in place.
Git-driven pipelines and platforms built to be handed over.
A pipeline your team drives — GitOps repo, build and release automation, runbooks, and the handover to go with them.
Models we design, train and run — on CPUs you already own.
A model and the pipeline around it — trained, exported for CPU inference, with the evaluation set it was measured on.
A custom computer-vision model, designed and trained from scratch — a lightweight CNN exported to ONNX for fast CPU inference — trained on roughly 200,000 labeled images across 5 cities in 3 countries.
It runs 6.8× faster than the baseline model it replaced, and it feeds a larger entity-resolution pipeline that combines geometric feature-matching and perceptual image hashing to confirm the same real-world item across independent data sources.
The pipeline runs in three stages, and each one discards candidates the next cannot recover.
Fig. 1 — the cascade. Each stage discards candidates; nothing downstream can recover them.
The lightweight CNN, trained from scratch and exported to ONNX, scores every image on ordinary CPUs — 1,147 images a second, no GPU in the loop.
Perceptual image hashing narrows the candidate field; geometric feature-matching then verifies the pairs that survive it.
The entity-resolution layer confirms the same real-world item across independent data sources, and hands the result to the surface that consumes it.
Four real incidents, root-caused and closed out — with the symptom that misled us left in.
A jumbo-frame MTU mismatch broke etcd's gRPC peer traffic — but ICMP tests kept passing, so the network looked fine.
Every connectivity check we had came back clean while etcd members kept losing quorum.
Root-caused via protocol-level reasoning: large gRPC frames were being silently dropped while small ping packets sailed through.
Recovered full quorum with zero data loss.
A CDN sitting in front of a tunnel-based origin broke silently, four different ways, before a fix stuck.
Each failure mode looked like a different problem, and each fix uncovered the next one.
Landed on a header-preserving CDN policy that keeps auth headers intact with zero redirects.
Documented as a reusable pattern for any CDN-in-front-of-tunnel setup.
Needed arm64 images built on amd64-only hardware, inside a cluster with a self-signed registry cert.
Chained QEMU emulation, BuildKit, and a registry-copy tool to get a clean multi-arch image out the other side.
Now a standing pipeline, not a one-off workaround.
Physically relocated a production Kubernetes cluster to a carrier datacenter with zero data loss.
Hit a storage-driver conflict on arrival that blocked pod scheduling.
Root-caused and resolved the same day — cluster back to full health before it mattered.
Every layer has at least two supported homes, and one of them is hardware you own.
Better scalability and cost efficiency
Increased reliability and faster delivery
Regulated identity and e-signature integrations, built to national compliance standards
Infrastructure built for regulated, high-availability financial systems
Global accessibility and scalability
Tell us what you are running. We will read it, come back within one working day, and if it is useful, spend thirty minutes reviewing your architecture — no charge, no obligation.
