You will get a self-hosted LLM infrastructure assessment for your workload
Project details
Most teams choose between self-hosted and API-based inference on instinct, then find out what it really costs six months later. This assessment answers the question before you commit hardware or sign up to a monthly bill.
I run production LLM inference on my own GPU nodes: llama.cpp built from source with custom CUDA binaries, quantization tuned per model and per card, served behind LiteLLM over OpenAI-compatible APIs. Not a hobby setup. It is the infrastructure I use daily for my own development work.
You tell me the workload, the concurrency, the latency you need and what data cannot leave your building. I tell you what hardware that actually requires, which model and quantization fits, what it costs against the API alternative over 12 and 36 months, and where the plan is likely to fall over.
Honest answers included. If a hosted API is genuinely the right call for your situation, the report will say so and explain why.
I run production LLM inference on my own GPU nodes: llama.cpp built from source with custom CUDA binaries, quantization tuned per model and per card, served behind LiteLLM over OpenAI-compatible APIs. Not a hobby setup. It is the infrastructure I use daily for my own development work.
You tell me the workload, the concurrency, the latency you need and what data cannot leave your building. I tell you what hardware that actually requires, which model and quantization fits, what it costs against the API alternative over 12 and 36 months, and where the plan is likely to fall over.
Honest answers included. If a hosted API is genuinely the right call for your situation, the report will say so and explain why.
AI Development Type
Deep Learning, Model TuningWhat's included
| Service Tiers |
Starter
$1,800
|
Standard
$3,600
|
Advanced
$8,100
|
|---|---|---|---|
| Delivery Time | 5 days | 12 days | 20 days |
Number of Revisions | 1 | 2 | 2 |
AI Model Integration | - | - | - |
Detailed Code Comments | - | - | - |
Knowledge Graph | - | - | - |
Model Documentation | - | - | - |
Ontology | - | - | - |
Source Code | - | - | - |
Taxonomy | - | - | - |
Frequently asked questions
About Niels
Game Developer & DevOps Engineer | Unity, .NET, Backend, AI & LLM
Leuven, Belgium - 7:23 pm local time
I get hired for the parts of a project that are hard to get right.
MULTIPLAYER AND NETCODE
Three shipped multiplayer titles on Photon (Bolt, Quantum, PUN). Authority handling, state replication, prediction and lag compensation, and the design compromises that make an online game feel fair. I have also retrofitted online co-op into a game that was never built for it, which is its own kind of education.
CONSOLE PORTING AND PLATFORM WORK
PlayStation, Xbox and Switch, certification included. Steamworks, Azure PlayFab, platform SDKs, and the store and account services around them.
BACKEND AND DEVOPS
Eighteen years of C#/.NET microservices in banking, logistics and HR, plus live-game backends and the whole delivery chain: self-hosted GitLab CE and Gitea, containerised build agents, CI/CD for Unity, Proxmox and Docker.
AI INFRASTRUCTURE AND TOOLING
I run production LLM inference on my own GPU nodes: llama.cpp built from source with custom CUDA binaries, quantization tuned per model, served behind LiteLLM over OpenAI-compatible APIs. I also built Welkin, an orchestration layer for multi-agent AI coding teams. If you need self-hosted inference, model serving, or AI wired into an existing product, that is daily work for me rather than a side interest.
I am most useful on projects where the brief is "this is broken and nobody knows why" or "we need this running on a platform it was never designed for". I write code meant to be read by whoever comes next, and I would rather tell you something takes three weeks than promise one.
Gameplay system, port, backend, build pipeline or AI integration: I bring engineering discipline to indie ambition.
Steps for completing your project
After purchasing the project, send requirements so Niels can start the project.
Delivery time starts when Niels receives requirements from you.
Niels works on your project following the steps below.
Revisions may occur after the delivery date.
Workload interview
A call and a short questionnaire: what you are running, expected concurrency, latency you need, what data cannot leave your infrastructure, and what budget you are working with.
Sizing and model selection
I work out what hardware the workload actually needs and which model and quantization fits it, rather than starting from whichever model is fashionable.