Stop bleeding compute capital.
Optimize your large algorithmic problems.

I architect Model-as-a-Service infrastructure and optimize low-level GPU workloads to slash your cloud bills. Fewer nodes. Higher throughput. Kernels that actually saturate the hardware you're already paying for.

▸ Request a Compute Audit cat ./services.md

///01 / the problem

Your GPUs are idle 80% of the time. You're paying full price for all of it.

Inefficient VRAM allocation, uncoalesced memory access, and Python-first wrappers turn a $40k/month compute budget into a runway-burning liability. Most teams scale horizontally — more instances, more nodes — instead of fixing what's already running.

×Kernels launching without saturating the SMs.
×Memory transfers hiding 3× more latency than the math.
×LLM serving stacks reserving VRAM they'll never touch.
×Autoscalers that scale spend, not throughput.

///02 / service catalog

Three engagements. One outcome: fewer dollars per FLOP.

S_01// service

MaaS Architecture Design

Architecting LLM workflow orchestrators, inference gateways, batching layers, and multi-tenant model routing built to scale without doubling your bill.

▸vLLM / TGI / Triton
▸Request batching
▸KV-cache strategy
▸Multi-tenant isolation

S_02// service

GPU Compute Profiling

Nsight Compute / Systems deep dives. I find the bottlenecks — memory coalescing, warp divergence, occupancy stalls — and ship a report with prioritized fixes.

▸Nsight Compute
▸Roofline analysis
▸Memory coalescing
▸Kernel timeline

S_03// service

CUDA / C++ Kernel Optimization

Replacing inefficient high-level wrappers with native performance. Custom CUDA kernels, C++ hot paths, and multi-GPU communication that actually uses NVLink.

▸Custom CUDA kernels
▸Tensor core usage
▸NCCL / NVLink
▸Fused ops

///03 / authority

$ whoami

role     : Independent HPC / GPU Engineer
research : M.Sc., HPC & Algorithms
lab      : University of Bergen (UiB)
focus    : Parallel algorithms
           Multi-GPU supercomputers
           Model-as-a-Service

Research-grade parallelism.
Production-grade delivery.

M.Sc. Researcher in HPC & Algorithms at the University of Bergen. Specializing in parallelizing algorithms for multi-GPU supercomputer environments — the same primitives that decide whether your inference cluster returns tokens in 40ms or 400ms.

I work with a small number of technical founders and CTOs per quarter. Every engagement starts with a fixed-scope Compute Audit — no retainers, no vendor lock-in.

// 04_contact

Ship faster kernels.
Spend less on GPUs.

Fixed scope, two-week Compute Audit. Deliverable: prioritized report with expected $/throughput impact per fix.

▸ Request a Compute Auditresponse < 48h · limited slots per quarter