inference optimization
Batching policy, KV cache layout and reuse, speculative decoding, quantization — worked end to end until tokens per second per GPU stops being the constraint.
gpu performance · agent reliability · ai infrastructure
We fix the AI you already run.
An engineering consultancy for teams running large models and agents in production. We profile what you have and report what we find, with the evidence attached.
What we do
gpu performance
When the framework has given everything it has, the time left is in the kernels. Warp-level profiling, fused attention paths, Triton, and PTX where the compiler will not cooperate.
ai infrastructure
Sharding that matches the hardware, autoscaling that understands token load rather than CPU, and the observability to prove the numbers hold under real traffic.
agent reliability
The evaluation set that decides whether it works, what one completed task actually costs once failed runs are counted, and what the system does on the day it is wrong.
What we ship
Two tools that came out of the consultancy. Both run on hardware you already own.
for developersattograd memory
Your agent forgets everything at midnight.
- free, MIT licensed
- one file, kept in your repo
- cursor, claude code, codex
for warehouseswarey
Cameras you only look at after it already went wrong.
- runs on your own box
- alert in under a second
- eight cameras, one machine
Two audits
Fixed price, fixed scope, fixed finish line. Both are the usual way in.
inference audit
For a serving stack that costs more than it should.
$2k – $5k5 – 10 days50% upfront
- Your stack profiled under production-shaped load
- Bottleneck report with the traces behind each claim
- Cost per million tokens now, and the reachable one
agent audit
For an agent already in production, and what each completed job costs.
$3k – $8k5 – 10 days50% upfront
- A labelled eval set built from your real runs, failures included
- Cost per successfully completed task, abandoned runs counted in
- The eval harness running in your CI, owned by you
How it runs
The same four stages whichever one you buy. Ten working days at the outside.
- day 0
scope
Whether there is enough headroom to be worth an audit, including when the answer is no. No fee.
- days 1–2
instrument
We reproduce your workload and profile it end to end. Nothing changes in production.
- days 3–5
diagnose
Findings ranked by what they cost you, evidence attached. The ones we were wrong about get reported too.
- days 6–10
hand back
The report, the cost model, and the changes in the order we would make them.
What we work in
If your stack is not here, say so — the profiling work transfers.
- vLLM
- TensorRT-LLM
- SGLang
- CUDA
- Triton
- PTX
- Nsight
- PyTorch
- Claude Agent SDK
- LangGraph
- MCP
- Langfuse
- OpenTelemetry
- Kubernetes
- AWS
- Terraform
When not to call us
- You have not shipped yet. There is nothing to profile until real traffic exists.
- You want somebody to own the system permanently. We measure, document, hand back, and leave.
- The job is to build the thing rather than examine it — that is our consultancy practice, and it runs differently.
- The spend is small enough that we would cost more than a year of the savings. We will say so on the first call.
Tell us what it costs you.
For a serving stack: the model, the GPUs, your traffic shape, and the number that is bothering you. For an agent: one run that went wrong and what should have happened instead. You will get a reply from the person who would do the work.