SRE & cost engineering for AI infrastructure
Velar is the SRE team behind your AI. We instrument your stack, cut your inference costs, and run your ML platform for you — with 7+ years of banking-grade reliability engineering and a serverless GPU cloud built from scratch.
7+ yrs SRE — banking & telco · ML platform ops on EKS (Flyte) · Grafana LGTM stack in production · built Velar Cloud
The problems we get called for.
“Our inference bill doubled in a quarter and nobody can say why.”
We instrument your stack and hand you a savings roadmap with numbers — token consumption doubled industry-wide last year; most of it is wasted.
“Our AI feature falls over every time we get rate-limited.”
60% of LLM production errors are rate limits. We build the retries, fallbacks and capacity planning that make it boring.
“We're paying for GPUs that run at 15% utilization.”
Most clusters run at 5–20% and nobody's watching. We make waste visible, then eliminate it.
“We've had an ML platform engineer role open for months.”
You can't hire this profile — almost nobody can. Get the senior half of that hire, fractional, this month.
Fixed scope, named deliverables. Pricing comes in a proposal after a 30-minute call.
Offer 01 · 2 weeks
We instrument your AI stack with Grafana — GPU utilization, $/token, cache hit rate, error and latency SLOs — and leave the dashboards running. You get a savings roadmap with numbers, whether you run on OpenAI's API or your own GPUs.
Offer 02 · 4–6 weeks
Your model serving production traffic on your cloud — vLLM, IaC, runbooks, a real handoff — with autoscaling, rate-limit handling and SLOs baked in. Infrastructure that survives launch day.
Offer 03 · monthly
The ML platform engineer you couldn't hire — as a monthly service. Deploys, scaling, monitoring, cost control. You get an endpoint and an invoice.
Every engagement starts with a 30-minute call. No lock-in — we leave you independent.
What we do — 01
GPU right-sizing, Modal vs. RunPod vs. self-hosted tradeoffs, and cost-per-token reduction across your serving stack. We find where the money leaks.
You get: a quantified savings plan — per workload, per GPU, per token.
What we do — 02
SLOs, incident response, and the full Grafana stack — utilization, latency, spend — applied to AI workloads. Banking-grade discipline for GPU infrastructure.
You get: dashboards, SLOs and runbooks your team actually uses.
What we do — 03
vLLM serving, persistent endpoints, and autoscaling — production-grade model deployment on your own infrastructure, without babysitting.
You get: an endpoint that survives launch day, on your cloud, documented.
What we do — 04
Flyte orchestration, batch pipelines, fine-tuning infrastructure — scheduling, isolation and utilization on shared clusters, like we ran in production.
You get: a platform your team can operate — scheduling, isolation, utilization.
What we take off your plate
You own the product. We own everything under it.
Before consulting, we designed, built, and operated a serverless GPU cloud — billing engine to scheduler. That product is the proof of work.
Velar Cloud is preparing its relaunch — join the waitlistMetering GPU jobs to the exact second, with scale-to-zero so idle time costs nothing.
Container images cached by content hash, cutting redeploys to under 15 seconds.
A single Python call spreading batch workloads across dozens of GPUs concurrently.
Persistent endpoints holding models hot in VRAM for zero cold-start inference.
redeploys via content-hash image caching
parallel fan-out from a single Python call
billing granularity — idle time costs zero
models held hot in VRAM by warm containers
04 — Contact
Thirty minutes is usually enough to spot the biggest cost or reliability win in your stack.
Or write to us directly: team@velar.run
Tell us what you're running — we reply within one business day.