Private AI from idea to production
LLMs and agentic harnesses, benchmarked and deployed on-premises, hybrid or on EU sovereign clouds.
LLMs and agentic harnesses, benchmarked and deployed on-premises, hybrid or on EU sovereign clouds.
Overview
Dielabs takes LLMs and agentic harnesses from use case to a working system. Alongside client work, we run a hands-on inference lab that benchmarks LLM serving and agentic workloads under production-like conditions: latency, throughput, capacity and cost per task. What we learn there becomes opinionated frameworks for day 0–1 decisions: what to deploy, where, and at what scale. Our lab work and products were selected to build with a European AI cloud:
What we do
From requirements to workload profile, SLOs and costs.
Workload modeling, capacity planning, SLO-driven GPU sizing. Deploying and tuning open-source models with vLLM, model parallelism and KV cache management.
Repeatable load tests to find bottlenecks and the performance/cost sweet spot. Monitoring TTFT, ITL, throughput, KV cache utilization and GPU saturation.
Agent harnesses, tool calling and MCP on self-hosted LLMs — and what multi-step, long-context agent traffic does to latency, KV cache reuse and cost per task.
RAG over company data, voice assistants and AI gateways, built on the serving stack — from proof of concept to a usable product.
Storage and virtualization (VMware, Nutanix, NetApp). RFI/RFQ, solution design and architecture validation.
Shortcuts
Working systems, documented end to end — architecture, decisions, problems solved and measured results.
See the works →17 LLM configurations, 50 coding exercises: quality, cost and time. One model in a league of its own, a tight cluster where price varies tenfold, and a 12 GB local node that reaches 70% — if you let it think.
Read the paper →About me — founder
15+ years in datacenter infrastructure (enterprise private cloud, Service Provider, Telco Cloud), in delivery and technical presales. First worked on AI infrastructure at Dell Technologies, contributing to the internal AI knowledge base and to early DGX deals.
Brings a datacenter design discipline to AI: a VCDX-style path from discovery and assessment to conceptual, logical and physical design, and the same “-ilities” (availability, manageability, performance, recoverability, security) applied to LLM serving and agentic systems.
Created Dielabs in late 2025 out of passion, believing that infrastructure and AI would end up speaking the same language. No playbooks existed, so the know-how and frameworks were built in a hands-on lab, where everything is tested before it reaches a client.
Bridges the gap between people’s expectations and what AI can really do today, replacing hype with numbers: workload profiles, SLOs, benchmarks and costs.
Open to Solution Architect & Presales roles, and to selected freelance projects.
