XARIV helps engineering teams reduce the time, cost, and uncertainty of building AI systems — from first workload sketch to procurement-ready architecture review.
No account required · Free to start
# define workload → get decision report
from xariv import LensClient
client = LensClient()
workload = client.workloads.define(
model="llama-3.1-70b",
traffic_rps=120,
slo_p99_ms=500,
)
report = workload.analyze()
report.bottleneck # memory_bandwidth
report.gpu_count # 31 (not 48)
report.recommendations # tp=2, fp8, batch=32
# benchmark with real traffic
pulse = client.pulse.run(
workload.id,
dataset="sharegpt-sample",
)
pulse.confirms(report) # p99 480ms @ 31 GPUscase study · h100 chat rollout
Spreadsheet estimate → XARIV model
GPU count
48 31
p99 latency
480ms
SLO 500ms met
capex saved
35%
38%
fewer GPUs provisioned
Right-sized a 70B MoE fleet before procurement
42%
p99 latency reduction
Identified memory-bandwidth bottleneck pre-deployment
$2.1M
annual infra savings
Avoided over-provisioned H100 cluster for chat workload
One workload. A different question each time.
Define a workload once in Lens. Swap models, scale traffic, or change hardware — get a decision report without rerunning production.
Swap to a smaller model?
model="llama-3.1-8b"Traffic doubles next quarter?
traffic_rps=240KV cache grows 2×?
context_len=32768Remove tensor parallelism?
tensor_parallel=1Move from H100 to L40S?
gpu="L40S"Sound familiar?
“My AI feature is slow, expensive, and I don't know why.”
Latency spikes, GPU bills, and no clear bottleneck.
“Leadership wants 10M users next quarter. Can we?”
Capacity questions with no model beyond spreadsheets.
“We have twelve tools and no single source of truth.”
Hugging Face, vLLM, Datadog, Excel — nothing connects the decision.
Powered by the control plane
Lens, Pulse, Atlas, Oracle, and Forge are modules inside one engineering control plane — not five unrelated products.
The decision workflow
Define, benchmark, analyze — complete in under 10 minutes. Calculators are optional step zero for quick ballpark checks.
Model, traffic, SLOs, and hardware constraints — one structured definition in Lens.
workload = lens.define(model="llama-70b")Start →Replay production or sample datasets in Pulse. Collect TTFT, throughput, and GPU telemetry.
pulse.run(workload.id, dataset="sharegpt")Start →Unified report with bottleneck explainability, ranked optimizations, and exportable decision.
report.export(format='pdf')Start →Replace the fragmented stack
Every step of the AI infrastructure lifecycle uses a different tool. XARIV owns the decision workflow between them.
Today — fragmented
Nobody owns the entire engineering decision workflow.
With XARIV — one platform
Enterprise lifecycle
Built for cross-functional teams
Product, ML, platform, SRE, finance, and leadership — same source of truth.
Platform in action
Global e-commerce platform. Planning a 32B chat model rollout on H100s with a 500ms p99 SLO — team estimated 48 GPUs from spreadsheet math. Lens predicted a memory-bandwidth bottleneck at 28 GPUs; Pulse replayed ShareGPT traffic and confirmed p99 at 31 GPUs with tensor parallelism. Provisioned 31 GPUs instead of 48 — 35% capex reduction with SLO met on day one.
“We stopped guessing GPU counts. XARIV Lens surfaced the memory-bandwidth constraint our spreadsheet model missed — and Pulse validated the fix before we cut a PO.”
From the blog
First-principles engineering analysis behind the platform models.
Hundreds of models, ten thousand GPUs, a nine-figure compute bill — and almost every design decision turns out to be one dial: how much do you pay, in money or complexity, to protect latency?
Hundreds of models, thousands of GPUs, a multi-million-dollar compute bill — and yet the scheduler dispatches about once a second. The hard part isn't throughput. It's trusting the number.
It looks read-heavy. It isn't. A per-user RAG at 220M-user scale is write-dominated, the LLM bill dwarfs the infrastructure, and one partitioning decision quietly solves four problems at once.
Start with a workload definition. Complete the workflow in under 10 minutes. Export when your team is ready to decide.